diff --git a/.gitignore b/.gitignore
index df46ba8..657d054 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,5 +1,6 @@
 .venv_cuda/
 __pycache__/
+debug_ocr/
 *.py[cod]
 .pytest_cache/
 .mypy_cache/
diff --git a/data/img_orig/air1.jpg b/data/img_orig/air1.jpg
deleted file mode 100644
index 8cc83f4..0000000
Binary files a/data/img_orig/air1.jpg and /dev/null differ
diff --git a/data/img_orig/air1.txt b/data/img_orig/air1.txt
deleted file mode 100644
index 143de00..0000000
--- a/data/img_orig/air1.txt
+++ /dev/null
@@ -1,4 +0,0 @@
-0 0.498437 0.479861 0.117188 0.168056
-0 0.000000 -0.138889 0.000000 0.000000
-0 0.000000 -0.138889 0.000000 0.000000
-0 0.000000 -0.138889 0.000000 0.000000
diff --git a/data/img_orig/air2.jpg b/data/img_orig/air2.jpg
deleted file mode 100644
index 231a6f8..0000000
Binary files a/data/img_orig/air2.jpg and /dev/null differ
diff --git a/data/img_orig/air2.txt b/data/img_orig/air2.txt
deleted file mode 100644
index 1ea565f..0000000
--- a/data/img_orig/air2.txt
+++ /dev/null
@@ -1 +0,0 @@
-0 0.602734 0.609028 0.110156 0.181944
diff --git a/data/img_orig/air3.jpg b/data/img_orig/air3.jpg
deleted file mode 100644
index eb5e5b4..0000000
Binary files a/data/img_orig/air3.jpg and /dev/null differ
diff --git a/data/img_orig/air3.txt b/data/img_orig/air3.txt
deleted file mode 100644
index fb1779b..0000000
--- a/data/img_orig/air3.txt
+++ /dev/null
@@ -1,2 +0,0 @@
-0 0.498828 0.459722 0.152344 0.191667
-0 0.530859 0.457639 0.214844 0.315278
diff --git a/data/img_orig/air4.jpg b/data/img_orig/air4.jpg
deleted file mode 100644
index 3abf362..0000000
Binary files a/data/img_orig/air4.jpg and /dev/null differ
diff --git a/data/img_orig/air4.txt b/data/img_orig/air4.txt
deleted file mode 100644
index 8f1fc44..0000000
--- a/data/img_orig/air4.txt
+++ /dev/null
@@ -1,9 +0,0 @@
-0 0.377734 0.401389 0.132031 0.155556
-0 0.529297 0.211111 0.144531 0.205556
-0 0.641406 0.393750 0.126563 0.143056
-0 0.744531 0.609028 0.148438 0.193056
-0 0.275000 0.642361 0.167187 0.206944
-0 0.526953 0.622917 0.160156 0.184722
-0 0.635156 0.807639 0.126563 0.134722
-0 0.375781 0.857639 0.126563 0.145833
-0 0.531641 0.949306 0.172656 0.101389
diff --git a/data/img_orig/air5.jpg b/data/img_orig/air5.jpg
deleted file mode 100644
index 1f21075..0000000
Binary files a/data/img_orig/air5.jpg and /dev/null differ
diff --git a/data/img_orig/air5.txt b/data/img_orig/air5.txt
deleted file mode 100644
index d4d8504..0000000
--- a/data/img_orig/air5.txt
+++ /dev/null
@@ -1 +0,0 @@
-0 0.462109 0.513889 0.186719 0.163889
diff --git a/data/img_orig/air6.jpg b/data/img_orig/air6.jpg
deleted file mode 100644
index a365ffb..0000000
Binary files a/data/img_orig/air6.jpg and /dev/null differ
diff --git a/data/img_orig/air6.txt b/data/img_orig/air6.txt
deleted file mode 100644
index 6645ee4..0000000
--- a/data/img_orig/air6.txt
+++ /dev/null
@@ -1,15 +0,0 @@
-0 0.140234 0.091667 0.244531 0.127778
-0 0.458594 0.102083 0.295312 0.115278
-0 0.792578 0.118750 0.272656 0.193056
-0 0.186328 0.302083 0.307031 0.154167
-0 0.498047 0.272222 0.261719 0.119444
-0 0.812500 0.316667 0.323438 0.158333
-0 0.203125 0.502778 0.367188 0.172222
-0 0.521094 0.487500 0.207813 0.255556
-0 0.798047 0.497222 0.303906 0.155556
-0 0.157031 0.714583 0.264062 0.140278
-0 0.480469 0.693056 0.321875 0.155556
-0 0.816406 0.688889 0.318750 0.133333
-0 0.182812 0.879167 0.314063 0.125000
-0 0.520312 0.882639 0.340625 0.181944
-0 0.832812 0.861111 0.260938 0.141667
diff --git a/data/img_orig/bird1.jpg b/data/img_orig/bird1.jpg
deleted file mode 100644
index 38d86b4..0000000
Binary files a/data/img_orig/bird1.jpg and /dev/null differ
diff --git a/data/img_orig/bird1.txt b/data/img_orig/bird1.txt
deleted file mode 100644
index c56538d..0000000
--- a/data/img_orig/bird1.txt
+++ /dev/null
@@ -1 +0,0 @@
-1 0.657812 0.411806 0.189063 0.223611
diff --git a/data/img_orig/bird2.jpg b/data/img_orig/bird2.jpg
deleted file mode 100644
index 0685783..0000000
Binary files a/data/img_orig/bird2.jpg and /dev/null differ
diff --git a/data/img_orig/bird2.txt b/data/img_orig/bird2.txt
deleted file mode 100644
index e08d6ca..0000000
--- a/data/img_orig/bird2.txt
+++ /dev/null
@@ -1,2 +0,0 @@
-1 0.552344 0.384028 0.159375 0.195833
-1 0.780078 0.595833 0.236719 0.133333
diff --git a/data/img_orig/bird3.jpg b/data/img_orig/bird3.jpg
deleted file mode 100644
index d011579..0000000
Binary files a/data/img_orig/bird3.jpg and /dev/null differ
diff --git a/data/img_orig/bird3.txt b/data/img_orig/bird3.txt
deleted file mode 100644
index 05eb5c5..0000000
--- a/data/img_orig/bird3.txt
+++ /dev/null
@@ -1,3 +0,0 @@
-1 0.508594 0.282639 0.107813 0.151389
-1 0.521875 0.470833 0.081250 0.100000
-1 0.527344 0.709722 0.162500 0.133333
diff --git a/data/img_orig/bird4.jpg b/data/img_orig/bird4.jpg
deleted file mode 100644
index f543ab5..0000000
Binary files a/data/img_orig/bird4.jpg and /dev/null differ
diff --git a/data/img_orig/bird4.txt b/data/img_orig/bird4.txt
deleted file mode 100644
index 3db7190..0000000
--- a/data/img_orig/bird4.txt
+++ /dev/null
@@ -1 +0,0 @@
-1 0.526953 0.502083 0.303906 0.226389
diff --git a/data/train.txt b/data/train.txt
index cfd322f..8ce6a24 100644
--- a/data/train.txt
+++ b/data/train.txt
@@ -1,201 +1,201 @@
-/home/abonvicini/darknetAB/darknet/data/img/thumb0100.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0101.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0102.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0103.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0104.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0105.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0106.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0107.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0108.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0109.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0110.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0111.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0112.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0113.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0114.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0115.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0116.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0117.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0118.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0119.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0120.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0121.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0122.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0123.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0124.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0125.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0126.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0127.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0128.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0129.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0130.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0131.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0132.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0133.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0134.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0135.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0136.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0137.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0138.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0139.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0140.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0141.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0142.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0143.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0144.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0145.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0146.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0147.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0148.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0149.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0150.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0151.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0152.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0153.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0154.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0155.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0156.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0157.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0158.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0159.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0160.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0161.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0162.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0163.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0164.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0165.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0166.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0167.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0168.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0169.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0170.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0171.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0172.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0173.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0174.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0175.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0176.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0177.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0178.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0179.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0180.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0181.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0182.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0183.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0184.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0185.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0186.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0187.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0188.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0189.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0190.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0191.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0192.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0193.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0194.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0195.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0196.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0197.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0198.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0199.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0200.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0201.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0202.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0203.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0204.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0205.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0206.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0207.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0208.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0209.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0210.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0211.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0212.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0213.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0214.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0215.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0216.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0217.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0218.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0219.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0220.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0221.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0222.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0223.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0224.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0225.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0226.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0227.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0228.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0229.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0230.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0231.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0232.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0233.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0234.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0235.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0236.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0237.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0238.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0239.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0240.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0241.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0242.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0243.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0244.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0245.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0246.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0247.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0248.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0249.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0250.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0251.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0252.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0253.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0254.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0255.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0256.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0257.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0258.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0259.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0260.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0261.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0262.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0263.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0264.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0265.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0266.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0267.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0268.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0269.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0270.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0271.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0272.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0273.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0274.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0275.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0276.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0277.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0278.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0279.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0280.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0281.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0282.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0283.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0284.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0285.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0286.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0287.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0288.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0289.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0290.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0291.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0292.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0293.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0294.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0295.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0296.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0297.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0298.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0299.jpg
-/home/abonvicini/darknetAB/darknet/data/img/thumb0300.jpg
+data/img/thumb0100.jpg
+data/img/thumb0101.jpg
+data/img/thumb0102.jpg
+data/img/thumb0103.jpg
+data/img/thumb0104.jpg
+data/img/thumb0105.jpg
+data/img/thumb0106.jpg
+data/img/thumb0107.jpg
+data/img/thumb0108.jpg
+data/img/thumb0109.jpg
+data/img/thumb0110.jpg
+data/img/thumb0111.jpg
+data/img/thumb0112.jpg
+data/img/thumb0113.jpg
+data/img/thumb0114.jpg
+data/img/thumb0115.jpg
+data/img/thumb0116.jpg
+data/img/thumb0117.jpg
+data/img/thumb0118.jpg
+data/img/thumb0119.jpg
+data/img/thumb0120.jpg
+data/img/thumb0121.jpg
+data/img/thumb0122.jpg
+data/img/thumb0123.jpg
+data/img/thumb0124.jpg
+data/img/thumb0125.jpg
+data/img/thumb0126.jpg
+data/img/thumb0127.jpg
+data/img/thumb0128.jpg
+data/img/thumb0129.jpg
+data/img/thumb0130.jpg
+data/img/thumb0131.jpg
+data/img/thumb0132.jpg
+data/img/thumb0133.jpg
+data/img/thumb0134.jpg
+data/img/thumb0135.jpg
+data/img/thumb0136.jpg
+data/img/thumb0137.jpg
+data/img/thumb0138.jpg
+data/img/thumb0139.jpg
+data/img/thumb0140.jpg
+data/img/thumb0141.jpg
+data/img/thumb0142.jpg
+data/img/thumb0143.jpg
+data/img/thumb0144.jpg
+data/img/thumb0145.jpg
+data/img/thumb0146.jpg
+data/img/thumb0147.jpg
+data/img/thumb0148.jpg
+data/img/thumb0149.jpg
+data/img/thumb0150.jpg
+data/img/thumb0151.jpg
+data/img/thumb0152.jpg
+data/img/thumb0153.jpg
+data/img/thumb0154.jpg
+data/img/thumb0155.jpg
+data/img/thumb0156.jpg
+data/img/thumb0157.jpg
+data/img/thumb0158.jpg
+data/img/thumb0159.jpg
+data/img/thumb0160.jpg
+data/img/thumb0161.jpg
+data/img/thumb0162.jpg
+data/img/thumb0163.jpg
+data/img/thumb0164.jpg
+data/img/thumb0165.jpg
+data/img/thumb0166.jpg
+data/img/thumb0167.jpg
+data/img/thumb0168.jpg
+data/img/thumb0169.jpg
+data/img/thumb0170.jpg
+data/img/thumb0171.jpg
+data/img/thumb0172.jpg
+data/img/thumb0173.jpg
+data/img/thumb0174.jpg
+data/img/thumb0175.jpg
+data/img/thumb0176.jpg
+data/img/thumb0177.jpg
+data/img/thumb0178.jpg
+data/img/thumb0179.jpg
+data/img/thumb0180.jpg
+data/img/thumb0181.jpg
+data/img/thumb0182.jpg
+data/img/thumb0183.jpg
+data/img/thumb0184.jpg
+data/img/thumb0185.jpg
+data/img/thumb0186.jpg
+data/img/thumb0187.jpg
+data/img/thumb0188.jpg
+data/img/thumb0189.jpg
+data/img/thumb0190.jpg
+data/img/thumb0191.jpg
+data/img/thumb0192.jpg
+data/img/thumb0193.jpg
+data/img/thumb0194.jpg
+data/img/thumb0195.jpg
+data/img/thumb0196.jpg
+data/img/thumb0197.jpg
+data/img/thumb0198.jpg
+data/img/thumb0199.jpg
+data/img/thumb0200.jpg
+data/img/thumb0201.jpg
+data/img/thumb0202.jpg
+data/img/thumb0203.jpg
+data/img/thumb0204.jpg
+data/img/thumb0205.jpg
+data/img/thumb0206.jpg
+data/img/thumb0207.jpg
+data/img/thumb0208.jpg
+data/img/thumb0209.jpg
+data/img/thumb0210.jpg
+data/img/thumb0211.jpg
+data/img/thumb0212.jpg
+data/img/thumb0213.jpg
+data/img/thumb0214.jpg
+data/img/thumb0215.jpg
+data/img/thumb0216.jpg
+data/img/thumb0217.jpg
+data/img/thumb0218.jpg
+data/img/thumb0219.jpg
+data/img/thumb0220.jpg
+data/img/thumb0221.jpg
+data/img/thumb0222.jpg
+data/img/thumb0223.jpg
+data/img/thumb0224.jpg
+data/img/thumb0225.jpg
+data/img/thumb0226.jpg
+data/img/thumb0227.jpg
+data/img/thumb0228.jpg
+data/img/thumb0229.jpg
+data/img/thumb0230.jpg
+data/img/thumb0231.jpg
+data/img/thumb0232.jpg
+data/img/thumb0233.jpg
+data/img/thumb0234.jpg
+data/img/thumb0235.jpg
+data/img/thumb0236.jpg
+data/img/thumb0237.jpg
+data/img/thumb0238.jpg
+data/img/thumb0239.jpg
+data/img/thumb0240.jpg
+data/img/thumb0241.jpg
+data/img/thumb0242.jpg
+data/img/thumb0243.jpg
+data/img/thumb0244.jpg
+data/img/thumb0245.jpg
+data/img/thumb0246.jpg
+data/img/thumb0247.jpg
+data/img/thumb0248.jpg
+data/img/thumb0249.jpg
+data/img/thumb0250.jpg
+data/img/thumb0251.jpg
+data/img/thumb0252.jpg
+data/img/thumb0253.jpg
+data/img/thumb0254.jpg
+data/img/thumb0255.jpg
+data/img/thumb0256.jpg
+data/img/thumb0257.jpg
+data/img/thumb0258.jpg
+data/img/thumb0259.jpg
+data/img/thumb0260.jpg
+data/img/thumb0261.jpg
+data/img/thumb0262.jpg
+data/img/thumb0263.jpg
+data/img/thumb0264.jpg
+data/img/thumb0265.jpg
+data/img/thumb0266.jpg
+data/img/thumb0267.jpg
+data/img/thumb0268.jpg
+data/img/thumb0269.jpg
+data/img/thumb0270.jpg
+data/img/thumb0271.jpg
+data/img/thumb0272.jpg
+data/img/thumb0273.jpg
+data/img/thumb0274.jpg
+data/img/thumb0275.jpg
+data/img/thumb0276.jpg
+data/img/thumb0277.jpg
+data/img/thumb0278.jpg
+data/img/thumb0279.jpg
+data/img/thumb0280.jpg
+data/img/thumb0281.jpg
+data/img/thumb0282.jpg
+data/img/thumb0283.jpg
+data/img/thumb0284.jpg
+data/img/thumb0285.jpg
+data/img/thumb0286.jpg
+data/img/thumb0287.jpg
+data/img/thumb0288.jpg
+data/img/thumb0289.jpg
+data/img/thumb0290.jpg
+data/img/thumb0291.jpg
+data/img/thumb0292.jpg
+data/img/thumb0293.jpg
+data/img/thumb0294.jpg
+data/img/thumb0295.jpg
+data/img/thumb0296.jpg
+data/img/thumb0297.jpg
+data/img/thumb0298.jpg
+data/img/thumb0299.jpg
+data/img/thumb0300.jpg
diff --git a/flywms.py b/flywms.py
new file mode 100644
index 0000000..b36f362
--- /dev/null
+++ b/flywms.py
@@ -0,0 +1,1644 @@
+import argparse
+import json
+import os
+import re
+import subprocess
+import sys
+import tempfile
+import threading
+import time
+from collections import deque
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+import cv2
+import numpy as np
+
+
+@dataclass(frozen=True)
+class FramePacket:
+    frame_id: int
+    timestamp: float
+    frame: np.ndarray
+    width: int
+    height: int
+    source: str
+
+
+@dataclass(frozen=True)
+class Detection:
+    class_id: int
+    class_name: str
+    confidence: float
+    bbox: tuple[int, int, int, int]
+
+
+@dataclass(frozen=True)
+class DetectionResult:
+    frame_id: int
+    timestamp: float
+    detections: list[Detection]
+    inference_ms: float
+    blob_ms: float
+    forward_ms: float
+    parse_ms: float
+    source_width: int
+    source_height: int
+
+
+@dataclass(frozen=True)
+class RoiPacket:
+    roi_id: int
+    source_frame_id: int
+    timestamp: float
+    class_name: str
+    confidence: float
+    bbox: tuple[int, int, int, int]
+    roi_image: np.ndarray
+    width: int
+    height: int
+
+
+@dataclass(frozen=True)
+class OcrResult:
+    roi_id: int
+    source_frame_id: int
+    timestamp: float
+    text: str
+    raw_text: str
+    bbox: tuple[int, int, int, int]
+    ocr_ms: float
+
+
+@dataclass(frozen=True)
+class OcrEngineResult:
+    digits: str
+    processed: np.ndarray
+    raw_text: str
+    code_roi: np.ndarray
+
+
+@dataclass
+class BufferStats:
+    pushed: int = 0
+    popped: int = 0
+    dropped_on_put: int = 0
+    skipped_on_pop: int = 0
+    waits: int = 0
+    max_depth_seen: int = 0
+
+
+@dataclass
+class RuntimeStats:
+    capture_frames: int = 0
+    display_frames: int = 0
+    yolo_submitted_frames: int = 0
+    quality_rejected_frames: int = 0
+    quality_cycles: int = 0
+    yolo_cycles: int = 0
+    ocr_cycles: int = 0
+    quality_total_ms: float = 0.0
+    quality_score_total: float = 0.0
+    yolo_total_ms: float = 0.0
+    yolo_blob_ms: float = 0.0
+    yolo_forward_ms: float = 0.0
+    yolo_parse_ms: float = 0.0
+    ocr_total_ms: float = 0.0
+    display_total_ms: float = 0.0
+    capture_read_total_ms: float = 0.0
+    last_capture_frame_id: int = 0
+    last_quality_score: float = 0.0
+    last_quality_passed: bool = True
+    last_yolo_frame_id: int = 0
+    last_ocr_frame_id: int = 0
+    last_detection_count: int = 0
+    last_detection_summary: str = ""
+
+
+class LatestBuffer:
+    def __init__(self, max_size: int, name: str):
+        if max_size < 1:
+            raise ValueError("max_size must be >= 1")
+        self.max_size = max_size
+        self.name = name
+        self._items: deque[Any] = deque(maxlen=max_size)
+        self._condition = threading.Condition()
+        self._stats = BufferStats()
+
+    def put(self, item: Any) -> None:
+        with self._condition:
+            if len(self._items) == self.max_size:
+                self._stats.dropped_on_put += 1
+            self._items.append(item)
+            self._stats.pushed += 1
+            self._stats.max_depth_seen = max(
+                self._stats.max_depth_seen,
+                len(self._items),
+            )
+            self._condition.notify_all()
+
+    def get_latest_blocking(self, stop_event: threading.Event) -> Any | None:
+        with self._condition:
+            while not self._items and not stop_event.is_set():
+                self._stats.waits += 1
+                self._condition.wait(timeout=0.1)
+
+            if stop_event.is_set():
+                return None
+
+            latest = self._items[-1]
+            skipped = len(self._items) - 1
+            self._stats.skipped_on_pop += skipped
+            self._items.clear()
+            self._stats.popped += 1
+            return latest
+
+    def wake_all(self) -> None:
+        with self._condition:
+            self._condition.notify_all()
+
+    def stats(self) -> BufferStats:
+        with self._condition:
+            return BufferStats(**self._stats.__dict__)
+
+    def depth(self) -> int:
+        with self._condition:
+            return len(self._items)
+
+
+class SharedState:
+    def __init__(self, ocr_history_size: int):
+        self._lock = threading.Lock()
+        self.latest_detection: DetectionResult | None = None
+        self.ocr_results: deque[OcrResult] = deque(maxlen=ocr_history_size)
+        self.debug_yolo_frame: np.ndarray | None = None
+        self.debug_ocr_frame: np.ndarray | None = None
+        self.debug_ocr_text: str = ""
+        self.stats = RuntimeStats()
+
+    def set_latest_detection(self, result: DetectionResult) -> None:
+        with self._lock:
+            self.latest_detection = result
+            self.stats.yolo_cycles += 1
+            self.stats.yolo_total_ms += result.inference_ms
+            self.stats.yolo_blob_ms += result.blob_ms
+            self.stats.yolo_forward_ms += result.forward_ms
+            self.stats.yolo_parse_ms += result.parse_ms
+            self.stats.last_yolo_frame_id = result.frame_id
+            self.stats.last_detection_count = len(result.detections)
+            counts: dict[str, int] = {}
+            for det in result.detections:
+                counts[det.class_name] = counts.get(det.class_name, 0) + 1
+            self.stats.last_detection_summary = ",".join(
+                f"{name}:{count}" for name, count in sorted(counts.items())
+            )
+
+    def get_latest_detection(self) -> DetectionResult | None:
+        with self._lock:
+            return self.latest_detection
+
+    def add_ocr_result(self, result: OcrResult) -> None:
+        with self._lock:
+            self.ocr_results.append(result)
+            self.stats.ocr_cycles += 1
+            self.stats.ocr_total_ms += result.ocr_ms
+            self.stats.last_ocr_frame_id = result.source_frame_id
+
+    def get_recent_ocr_results(self) -> list[OcrResult]:
+        with self._lock:
+            return list(self.ocr_results)
+
+    def set_debug_yolo_frame(self, frame: np.ndarray | None) -> None:
+        with self._lock:
+            self.debug_yolo_frame = None if frame is None else frame.copy()
+
+    def set_debug_ocr_frame(self, frame: np.ndarray | None, text: str = "") -> None:
+        with self._lock:
+            self.debug_ocr_frame = None if frame is None else frame.copy()
+            self.debug_ocr_text = text
+
+    def get_debug_frames(self) -> tuple[np.ndarray | None, np.ndarray | None, str]:
+        with self._lock:
+            yolo = None if self.debug_yolo_frame is None else self.debug_yolo_frame.copy()
+            ocr = None if self.debug_ocr_frame is None else self.debug_ocr_frame.copy()
+            return yolo, ocr, self.debug_ocr_text
+
+    def add_capture_read(self, frame_id: int, read_ms: float) -> None:
+        with self._lock:
+            self.stats.capture_frames += 1
+            self.stats.capture_read_total_ms += read_ms
+            self.stats.last_capture_frame_id = frame_id
+
+    def add_quality_result(
+        self,
+        score: float,
+        passed: bool,
+        elapsed_ms: float,
+        submitted_to_yolo: bool,
+    ) -> None:
+        with self._lock:
+            self.stats.quality_cycles += 1
+            self.stats.quality_score_total += score
+            self.stats.quality_total_ms += elapsed_ms
+            self.stats.last_quality_score = score
+            self.stats.last_quality_passed = passed
+            if submitted_to_yolo:
+                self.stats.yolo_submitted_frames += 1
+            else:
+                self.stats.quality_rejected_frames += 1
+
+    def add_display(self, display_ms: float) -> None:
+        with self._lock:
+            self.stats.display_frames += 1
+            self.stats.display_total_ms += display_ms
+
+    def snapshot_stats(self) -> RuntimeStats:
+        with self._lock:
+            return RuntimeStats(**self.stats.__dict__)
+
+
+class IdGenerator:
+    def __init__(self, start: int = 1):
+        self._value = start
+        self._lock = threading.Lock()
+
+    def next(self) -> int:
+        with self._lock:
+            value = self._value
+            self._value += 1
+            return value
+
+
+def parse_args():
+    ap = argparse.ArgumentParser()
+
+    ap.add_argument("-v", "--video", default=None,
+                    help="Percorso video. Se omesso usa webcam 0")
+    ap.add_argument("--weights", default="yolov2.weights",
+                    help="File pesi YOLOv2")
+    ap.add_argument("--config", default="yolov2.cfg",
+                    help="File config YOLOv2")
+    ap.add_argument("--labels", default="labels.txt",
+                    help="File labels classi")
+    ap.add_argument("--tesseract-cmd", default=None,
+                    help="Percorso esplicito a tesseract.exe")
+
+    ap.add_argument("--backend", choices=["cpu", "cuda", "cuda-fp16"],
+                    default="cpu", help="Backend OpenCV DNN")
+    ap.add_argument("--input-size", type=int, default=416,
+                    help="Dimensione input YOLO")
+    ap.add_argument("--swap-rb", action="store_true",
+                    help="Scambia canali R/B nella blob YOLO")
+    ap.add_argument("--frame-buffer-size", type=int, default=10,
+                    help="Dimensione latest buffer frame")
+    ap.add_argument("--roi-buffer-size", type=int, default=20,
+                    help="Dimensione latest buffer ROI")
+    ap.add_argument("--ocr-history-size", type=int, default=100,
+                    help="Numero risultati OCR recenti mantenuti in memoria")
+    ap.add_argument("--preview-width", type=int, default=1280,
+                    help="Larghezza massima preview")
+    ap.add_argument("--stats-interval", type=float, default=2.0,
+                    help="Secondi tra riepiloghi prestazioni")
+    ap.add_argument("--max-frames", type=int, default=0,
+                    help="Numero massimo frame da leggere; 0 = fino a fine stream")
+    ap.add_argument("--drain-seconds", type=float, default=0.0,
+                    help="Secondi di attesa dopo max/fine stream per benchmark headless")
+    ap.add_argument("--realtime-playback", action="store_true",
+                    help="Per file video, limita il loop al framerate del video")
+    ap.add_argument("--opencv-threads", type=int, default=1,
+                    help="Numero thread OpenCV")
+    ap.add_argument("--quality-filter", action="store_true",
+                    help="Filtra i frame troppo sfocati/mossi prima del buffer YOLO")
+    ap.add_argument("--blur-metric", choices=["laplacian", "tenengrad"],
+                    default="laplacian", help="Metrica nitidezza usata dal filtro qualita'")
+    ap.add_argument("--min-sharpness", type=float, default=80.0,
+                    help="Soglia minima nitidezza per inviare il frame a YOLO")
+    ap.add_argument("--blur-resize-width", type=int, default=320,
+                    help="Larghezza usata per ridurre il frame prima della metrica blur")
+    ap.add_argument("--debug-quality-log", action="store_true",
+                    help="Logga lo score qualita' di ogni frame")
+    ap.add_argument("--debug-rejected-window", action="store_true",
+                    help="Mostra una finestra con i frame scartati dal filtro qualita'")
+
+    ap.add_argument("--min-confidence", type=float, default=0.30,
+                    help="Soglia minima confidenza")
+    ap.add_argument("--nms-threshold", type=float, default=0.40,
+                    help="Soglia NMS")
+    ap.add_argument("--use-nms", action="store_true",
+                    help="Applica NMS alle detection; default off per compatibilita' YOLOv2")
+    ap.add_argument("--label-class", default="etichetta",
+                    help="Nome classe etichetta su cui fare OCR")
+    ap.add_argument("--min-label-width", type=int, default=50,
+                    help="Larghezza minima bbox etichetta")
+    ap.add_argument("--min-label-height", type=int, default=20,
+                    help="Altezza minima bbox etichetta")
+    ap.add_argument("--max-roi-per-frame", type=int, default=2,
+                    help="Numero massimo ROI etichetta inviate a OCR per detection")
+    ap.add_argument("--infer-gaylord-from-label", action="store_true",
+                    help="Disegna un box gaylord stimato partendo dalle etichette se YOLO non trova gaylord")
+    ap.add_argument("--inferred-gaylord-width-factor", type=float, default=3.6,
+                    help="Larghezza box gaylord stimato rispetto alla label")
+    ap.add_argument("--inferred-gaylord-height-factor", type=float, default=4.2,
+                    help="Altezza box gaylord stimato rispetto alla label")
+    ap.add_argument("--inferred-gaylord-y-shift", type=float, default=1.35,
+                    help="Spostamento verticale verso il basso, in multipli dell'altezza label")
+    ap.add_argument("--slot-size", type=int, default=120,
+                    help="Dimensione griglia per cooldown OCR")
+    ap.add_argument("--ocr-cooldown-sec", type=float, default=1.0,
+                    help="Secondi minimi prima di reinviare OCR sulla stessa zona")
+    ap.add_argument("--ocr-min-digits", type=int, default=2,
+                    help="Numero minimo cifre per lettura valida")
+    ap.add_argument("--ocr-backend", choices=["tesseract", "paddle", "easyocr"],
+                    default="paddle", help="Motore OCR da usare")
+    ap.add_argument("--ocr-lang", default="en",
+                    help="Lingua OCR")
+    ap.add_argument("--easyocr-gpu", choices=["auto", "on", "off"],
+                    default="auto", help="Uso GPU EasyOCR")
+    ap.add_argument("--easyocr-mode", choices=["subprocess", "inprocess"],
+                    default="subprocess", help="Modalita' EasyOCR")
+    ap.add_argument("--easyocr-worker", action="store_true",
+                    help=argparse.SUPPRESS)
+    ap.add_argument("--ocr-input", choices=["roi", "processed"],
+                    default="roi", help="Immagine passata al motore OCR")
+    ap.add_argument("--ocr-code-mode", choices=["full", "fixed-band", "large-components"],
+                    default="fixed-band", help="Prefiltro per isolare il codice grande")
+    ap.add_argument("--ocr-scale", type=float, default=1.5,
+                    help="Fattore di ingrandimento preprocess OCR")
+    ap.add_argument("--ocr-max-width", type=int, default=900,
+                    help="Larghezza massima immagine inviata all'OCR")
+    ap.add_argument("--ocr-band-x1", type=float, default=0.0,
+                    help="Crop fisso OCR: x iniziale percentuale 0..1")
+    ap.add_argument("--ocr-band-y1", type=float, default=0.0,
+                    help="Crop fisso OCR: y iniziale percentuale 0..1")
+    ap.add_argument("--ocr-band-x2", type=float, default=1.0,
+                    help="Crop fisso OCR: x finale percentuale 0..1")
+    ap.add_argument("--ocr-band-y2", type=float, default=1.0,
+                    help="Crop fisso OCR: y finale percentuale 0..1")
+    ap.add_argument("--ocr-component-min-height-ratio", type=float, default=0.22,
+                    help="Altezza minima componente grande rispetto alla ROI")
+    ap.add_argument("--ocr-component-min-area-ratio", type=float, default=0.002,
+                    help="Area minima componente grande rispetto alla ROI")
+    ap.add_argument("--ocr-component-pad-ratio", type=float, default=0.08,
+                    help="Padding crop finale componenti grandi")
+    ap.add_argument("--ocr-pad-ratio", type=float, default=0.20,
+                    help="Padding bbox etichetta prima dell'OCR")
+    ap.add_argument("--ocr-submit-min-interval", type=float, default=2.0,
+                    help="Secondi minimi globali tra due ROI inviate all'OCR")
+    ap.add_argument("--ocr-max-pending", type=int, default=1,
+                    help="Numero massimo ROI pendenti prima di saltare nuovi invii OCR")
+    ap.add_argument("--paddle-text-det-limit-side-len", type=int, default=320,
+                    help="Parametro PaddleOCR text_det_limit_side_len")
+    ap.add_argument("--paddle-text-rec-score-thresh", type=float, default=0.0,
+                    help="Soglia riconoscimento PaddleOCR")
+    ap.add_argument("--print-all-ocr", action="store_true",
+                    help="Stampa anche OCR grezzi non validi")
+    ap.add_argument("--save-ocr-roi-dir", default=None,
+                    help="Directory dove salvare ROI OCR raw/code/processed per debug")
+    ap.add_argument("--no-ocr", action="store_true",
+                    help="Disabilita OCR; utile per benchmark YOLO/capture")
+
+    ap.add_argument("--debug-yolo-window", action="store_true",
+                    help="Mostra una finestra debug con l'ultimo frame YOLO")
+    ap.add_argument("--debug-ocr-window", action="store_true",
+                    help="Mostra una finestra debug con l'ultima ROI preprocessata")
+    ap.add_argument("--debug-yolo-output", action="store_true",
+                    help="Logga shape e confidenze grezze dell'output YOLO")
+    ap.add_argument("--debug-yolo-top", type=int, default=0,
+                    help="Logga le top N righe raw YOLO per confidenza classe")
+    ap.add_argument("--debug-inferred-gaylord", action="store_true",
+                    help="Logga quanti box gaylord stimati vengono generati")
+    ap.add_argument("--no-display", action="store_true",
+                    help="Disabilita finestre video, utile per benchmark")
+
+    return ap.parse_args()
+
+
+def log(msg: str) -> None:
+    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)
+
+
+def require_file(path_str: str, description: str) -> Path:
+    path = Path(path_str)
+    if not path.exists():
+        log(f"ERRORE: {description} non trovato: {path}")
+        sys.exit(1)
+    return path
+
+
+def load_classes(labels_path: str) -> list[str]:
+    with open(labels_path, "rt", encoding="utf-8") as f:
+        classes = [line.strip() for line in f if line.strip()]
+    if not classes:
+        log("ERRORE: labels.txt vuoto")
+        sys.exit(1)
+    return classes
+
+
+def open_capture(video_arg: str | None):
+    if video_arg is None:
+        cap = cv2.VideoCapture(0, cv2.CAP_DSHOW)
+        if not cap.isOpened():
+            cap = cv2.VideoCapture(0)
+        return cap, "camera:0"
+
+    if str(video_arg).isdigit():
+        idx = int(video_arg)
+        cap = cv2.VideoCapture(idx, cv2.CAP_DSHOW)
+        if not cap.isOpened():
+            cap = cv2.VideoCapture(idx)
+        return cap, f"camera:{idx}"
+
+    return cv2.VideoCapture(video_arg), str(video_arg)
+
+
+def configure_net_backend(net, backend: str) -> None:
+    if backend == "cpu":
+        net.setPreferableBackend(cv2.dnn.DNN_BACKEND_OPENCV)
+        net.setPreferableTarget(cv2.dnn.DNN_TARGET_CPU)
+        return
+
+    net.setPreferableBackend(cv2.dnn.DNN_BACKEND_CUDA)
+    if backend == "cuda-fp16":
+        net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA_FP16)
+    else:
+        net.setPreferableTarget(cv2.dnn.DNN_TARGET_CUDA)
+
+
+def resize_preview(frame: np.ndarray, max_width: int) -> np.ndarray:
+    h, w = frame.shape[:2]
+    if max_width <= 0 or w <= max_width:
+        return frame
+    scale = max_width / float(w)
+    new_w = int(w * scale)
+    new_h = int(h * scale)
+    return cv2.resize(frame, (new_w, new_h), interpolation=cv2.INTER_LINEAR)
+
+
+def clip_box(x1: int, y1: int, x2: int, y2: int,
+             w: int, h: int) -> tuple[int, int, int, int]:
+    x1 = max(0, min(x1, w - 1))
+    y1 = max(0, min(y1, h - 1))
+    x2 = max(0, min(x2, w - 1))
+    y2 = max(0, min(y2, h - 1))
+    return x1, y1, x2, y2
+
+
+def expand_box(x1: int, y1: int, x2: int, y2: int,
+               frame_w: int, frame_h: int,
+               pad_ratio: float = 0.08) -> tuple[int, int, int, int]:
+    bw = x2 - x1
+    bh = y2 - y1
+    pad_x = int(bw * pad_ratio)
+    pad_y = int(bh * pad_ratio)
+    return clip_box(
+        x1 - pad_x,
+        y1 - pad_y,
+        x2 + pad_x,
+        y2 + pad_y,
+        frame_w,
+        frame_h,
+    )
+
+
+def quantized_slot_key(bbox: tuple[int, int, int, int], slot_size: int) -> tuple[int, int]:
+    x1, y1, x2, y2 = bbox
+    cx = (x1 + x2) // 2
+    cy = (y1 + y2) // 2
+    return cx // slot_size, cy // slot_size
+
+
+def limit_width(image: np.ndarray, max_width: int) -> np.ndarray:
+    if max_width <= 0:
+        return image
+    h, w = image.shape[:2]
+    if w <= max_width:
+        return image
+    scale = max_width / float(w)
+    new_h = max(1, int(h * scale))
+    return cv2.resize(image, (max_width, new_h), interpolation=cv2.INTER_AREA)
+
+
+def preprocess_for_ocr(roi: np.ndarray, scale: float = 1.5, max_width: int = 900) -> np.ndarray:
+    gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
+    if scale != 1.0:
+        gray = cv2.resize(gray, None, fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC)
+    gray = cv2.GaussianBlur(gray, (3, 3), 0)
+    gray = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY + cv2.THRESH_OTSU)[1]
+    gray = cv2.copyMakeBorder(
+        gray, 8, 8, 8, 8,
+        borderType=cv2.BORDER_CONSTANT,
+        value=255,
+    )
+    gray = limit_width(gray, max_width)
+    return gray
+
+
+def crop_fixed_band(
+    roi: np.ndarray,
+    x1_ratio: float,
+    y1_ratio: float,
+    x2_ratio: float,
+    y2_ratio: float,
+) -> np.ndarray:
+    h, w = roi.shape[:2]
+    x1 = int(max(0.0, min(1.0, x1_ratio)) * w)
+    y1 = int(max(0.0, min(1.0, y1_ratio)) * h)
+    x2 = int(max(0.0, min(1.0, x2_ratio)) * w)
+    y2 = int(max(0.0, min(1.0, y2_ratio)) * h)
+    x1, y1, x2, y2 = clip_box(x1, y1, x2, y2, w, h)
+    if x2 <= x1 or y2 <= y1:
+        return roi
+    return roi[y1:y2, x1:x2].copy()
+
+
+def crop_large_components(
+    roi: np.ndarray,
+    min_height_ratio: float,
+    min_area_ratio: float,
+    pad_ratio: float,
+) -> np.ndarray:
+    h, w = roi.shape[:2]
+    if h <= 0 or w <= 0:
+        return roi
+
+    gray = cv2.cvtColor(roi, cv2.COLOR_BGR2GRAY)
+    gray = cv2.GaussianBlur(gray, (3, 3), 0)
+    binary = cv2.threshold(
+        gray,
+        0,
+        255,
+        cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU,
+    )[1]
+
+    kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (3, 3))
+    binary = cv2.morphologyEx(binary, cv2.MORPH_OPEN, kernel, iterations=1)
+
+    contours, _ = cv2.findContours(binary, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    min_h = max(3, int(h * min_height_ratio))
+    min_area = max(4, int(w * h * min_area_ratio))
+
+    boxes: list[tuple[int, int, int, int]] = []
+    for contour in contours:
+        x, y, bw, bh = cv2.boundingRect(contour)
+        area = bw * bh
+        if bh < min_h or area < min_area:
+            continue
+        aspect = bw / float(max(1, bh))
+        if aspect < 0.12 or aspect > 1.25:
+            continue
+        boxes.append((x, y, x + bw, y + bh))
+
+    if not boxes:
+        return roi
+
+    x1 = min(box[0] for box in boxes)
+    y1 = min(box[1] for box in boxes)
+    x2 = max(box[2] for box in boxes)
+    y2 = max(box[3] for box in boxes)
+    x1, y1, x2, y2 = expand_box(x1, y1, x2, y2, w, h, pad_ratio=pad_ratio)
+    if x2 <= x1 or y2 <= y1:
+        return roi
+    return roi[y1:y2, x1:x2].copy()
+
+
+def extract_code_roi(roi: np.ndarray, args) -> np.ndarray:
+    if args.ocr_code_mode == "full":
+        return roi
+    if args.ocr_code_mode == "fixed-band":
+        return crop_fixed_band(
+            roi,
+            args.ocr_band_x1,
+            args.ocr_band_y1,
+            args.ocr_band_x2,
+            args.ocr_band_y2,
+        )
+    return crop_large_components(
+        roi,
+        args.ocr_component_min_height_ratio,
+        args.ocr_component_min_area_ratio,
+        args.ocr_component_pad_ratio,
+    )
+
+
+def ocr_digits_only(roi: np.ndarray, pytesseract_module: Any) -> OcrEngineResult:
+    processed = preprocess_for_ocr(roi)
+    config = r"--oem 3 --psm 7 -c tessedit_char_whitelist=0123456789"
+    raw_text = pytesseract_module.image_to_string(processed, config=config)
+    digits = re.sub(r"\D+", "", raw_text)
+    return OcrEngineResult(digits, processed, raw_text, roi)
+
+
+class TesseractOcrEngine:
+    def __init__(self, args):
+        import pytesseract
+        if args.tesseract_cmd:
+            pytesseract.pytesseract.tesseract_cmd = args.tesseract_cmd
+        self._pytesseract = pytesseract
+        self._args = args
+
+    def read_digits(self, roi: np.ndarray) -> OcrEngineResult:
+        code_roi = extract_code_roi(roi, self._args)
+        return ocr_digits_only(code_roi, self._pytesseract)
+
+
+class PaddleOcrEngine:
+    def __init__(self, args):
+        from paddleocr import PaddleOCR
+        self._ocr = PaddleOCR(
+            lang=args.ocr_lang,
+            use_doc_orientation_classify=False,
+            use_doc_unwarping=False,
+            use_textline_orientation=False,
+            text_det_limit_side_len=args.paddle_text_det_limit_side_len,
+            text_rec_score_thresh=args.paddle_text_rec_score_thresh,
+        )
+        self._input_mode = args.ocr_input
+        self._args = args
+
+    def read_digits(self, roi: np.ndarray) -> OcrEngineResult:
+        code_roi = extract_code_roi(roi, self._args)
+        code_roi = limit_width(code_roi, self._args.ocr_max_width)
+        processed = preprocess_for_ocr(
+            code_roi,
+            scale=self._args.ocr_scale,
+            max_width=self._args.ocr_max_width,
+        )
+        ocr_input = processed if self._input_mode == "processed" else code_roi
+        result = self._ocr.predict(ocr_input)
+        texts: list[str] = []
+        for item in result:
+            if isinstance(item, dict):
+                rec_texts = item.get("rec_texts") or item.get("texts") or []
+                if isinstance(rec_texts, str):
+                    texts.append(rec_texts)
+                else:
+                    texts.extend(str(text) for text in rec_texts)
+            elif isinstance(item, (list, tuple)):
+                texts.append(str(item))
+
+        raw_text = " ".join(texts)
+        digits = re.sub(r"\D+", "", raw_text)
+        return OcrEngineResult(digits, processed, raw_text, code_roi)
+
+
+class EasyOcrInProcessEngine:
+    def __init__(self, args):
+        import easyocr
+        self._args = args
+        self._input_mode = args.ocr_input
+        langs = [part.strip() for part in args.ocr_lang.split(",") if part.strip()]
+        if not langs:
+            langs = ["en"]
+
+        requested_gpu = args.easyocr_gpu != "off"
+        if args.easyocr_gpu == "on":
+            self._reader = easyocr.Reader(langs, gpu=True, verbose=False)
+            self._using_gpu = True
+        elif requested_gpu:
+            try:
+                self._reader = easyocr.Reader(langs, gpu=True, verbose=False)
+                self._using_gpu = True
+            except Exception as exc:
+                log(f"EasyOCR GPU non disponibile, fallback CPU: {exc}")
+                self._reader = easyocr.Reader(langs, gpu=False, verbose=False)
+                self._using_gpu = False
+        else:
+            self._reader = easyocr.Reader(langs, gpu=False, verbose=False)
+            self._using_gpu = False
+
+        log(f"EasyOCR device: {'gpu' if self._using_gpu else 'cpu'}")
+
+    def read_digits(self, roi: np.ndarray) -> OcrEngineResult:
+        code_roi = extract_code_roi(roi, self._args)
+        code_roi = limit_width(code_roi, self._args.ocr_max_width)
+        processed = preprocess_for_ocr(
+            code_roi,
+            scale=self._args.ocr_scale,
+            max_width=self._args.ocr_max_width,
+        )
+        ocr_input = processed if self._input_mode == "processed" else code_roi
+        result = self._reader.readtext(
+            ocr_input,
+            allowlist="0123456789",
+            detail=1,
+            paragraph=False,
+        )
+        texts: list[str] = []
+        for item in result:
+            if isinstance(item, (list, tuple)) and len(item) >= 2:
+                texts.append(str(item[1]))
+            else:
+                texts.append(str(item))
+
+        raw_text = " ".join(texts)
+        digits = re.sub(r"\D+", "", raw_text)
+        return OcrEngineResult(digits, processed, raw_text, code_roi)
+
+
+class EasyOcrProcessEngine:
+    def __init__(self, args):
+        self._args = args
+        self._input_mode = args.ocr_input
+        self._tmpdir = tempfile.TemporaryDirectory(prefix="flywms_easyocr_")
+        cmd = [
+            sys.executable,
+            str(Path(__file__).resolve()),
+            "--easyocr-worker",
+            "--ocr-lang", args.ocr_lang,
+            "--easyocr-gpu", args.easyocr_gpu,
+        ]
+        env = os.environ.copy()
+        env["PYTHONUTF8"] = "1"
+        self._proc = subprocess.Popen(
+            cmd,
+            stdin=subprocess.PIPE,
+            stdout=subprocess.PIPE,
+            stderr=subprocess.PIPE,
+            text=True,
+            encoding="utf-8",
+            errors="replace",
+            env=env,
+        )
+        ready = self._read_json_line()
+        if ready.get("event") != "ready":
+            raise RuntimeError(f"EasyOCR worker non pronto: {ready}")
+        log(f"EasyOCR worker process device: {ready.get('device', 'unknown')}")
+
+    def _read_json_line(self) -> dict[str, Any]:
+        assert self._proc.stdout is not None
+        while True:
+            line = self._proc.stdout.readline()
+            if line == "":
+                err = ""
+                if self._proc.stderr is not None:
+                    err = self._proc.stderr.read()
+                raise RuntimeError(f"EasyOCR worker terminato: {err.strip()}")
+            line = line.strip()
+            if not line:
+                continue
+            try:
+                return json.loads(line)
+            except json.JSONDecodeError:
+                continue
+
+    def read_digits(self, roi: np.ndarray) -> OcrEngineResult:
+        if self._proc.poll() is not None:
+            raise RuntimeError("EasyOCR worker non attivo")
+
+        code_roi = extract_code_roi(roi, self._args)
+        code_roi = limit_width(code_roi, self._args.ocr_max_width)
+        processed = preprocess_for_ocr(
+            code_roi,
+            scale=self._args.ocr_scale,
+            max_width=self._args.ocr_max_width,
+        )
+        ocr_input = processed if self._input_mode == "processed" else code_roi
+        image_path = Path(self._tmpdir.name) / f"ocr_{time.perf_counter_ns()}.png"
+        cv2.imwrite(str(image_path), ocr_input)
+
+        assert self._proc.stdin is not None
+        self._proc.stdin.write(json.dumps({"path": str(image_path)}) + "\n")
+        self._proc.stdin.flush()
+        response = self._read_json_line()
+        try:
+            image_path.unlink(missing_ok=True)
+        except OSError:
+            pass
+
+        if response.get("error"):
+            raise RuntimeError(str(response["error"]))
+        raw_text = str(response.get("raw_text", ""))
+        digits = re.sub(r"\D+", "", raw_text)
+        return OcrEngineResult(digits, processed, raw_text, code_roi)
+
+    def close(self) -> None:
+        if getattr(self, "_proc", None) is None:
+            return
+        if self._proc.poll() is None:
+            try:
+                assert self._proc.stdin is not None
+                self._proc.stdin.write(json.dumps({"cmd": "stop"}) + "\n")
+                self._proc.stdin.flush()
+            except Exception:
+                pass
+            try:
+                self._proc.wait(timeout=3)
+            except subprocess.TimeoutExpired:
+                self._proc.kill()
+        self._tmpdir.cleanup()
+
+    def __del__(self):
+        try:
+            self.close()
+        except Exception:
+            pass
+
+
+def create_ocr_engine(args):
+    if args.ocr_backend == "paddle":
+        return PaddleOcrEngine(args)
+    if args.ocr_backend == "easyocr":
+        if args.easyocr_mode == "inprocess":
+            return EasyOcrInProcessEngine(args)
+        return EasyOcrProcessEngine(args)
+    return TesseractOcrEngine(args)
+
+
+def run_easyocr_worker(args) -> int:
+    import easyocr
+
+    langs = [part.strip() for part in args.ocr_lang.split(",") if part.strip()] or ["en"]
+    requested_gpu = args.easyocr_gpu != "off"
+    using_gpu = False
+    if args.easyocr_gpu == "on":
+        reader = easyocr.Reader(langs, gpu=True, verbose=False)
+        using_gpu = True
+    elif requested_gpu:
+        try:
+            reader = easyocr.Reader(langs, gpu=True, verbose=False)
+            using_gpu = True
+        except Exception:
+            reader = easyocr.Reader(langs, gpu=False, verbose=False)
+    else:
+        reader = easyocr.Reader(langs, gpu=False, verbose=False)
+
+    print(json.dumps({"event": "ready", "device": "gpu" if using_gpu else "cpu"}), flush=True)
+    for line in sys.stdin:
+        try:
+            request = json.loads(line)
+            if request.get("cmd") == "stop":
+                break
+            result = reader.readtext(
+                request["path"],
+                allowlist="0123456789",
+                detail=1,
+                paragraph=False,
+            )
+            texts: list[str] = []
+            for item in result:
+                if isinstance(item, (list, tuple)) and len(item) >= 2:
+                    texts.append(str(item[1]))
+                else:
+                    texts.append(str(item))
+            raw_text = " ".join(texts)
+            print(json.dumps({"raw_text": raw_text}), flush=True)
+        except Exception as exc:
+            print(json.dumps({"error": str(exc)}), flush=True)
+    return 0
+
+
+def save_ocr_debug_images(
+    output_dir: Path,
+    roi_packet: RoiPacket,
+    engine_result: OcrEngineResult,
+) -> None:
+    output_dir.mkdir(parents=True, exist_ok=True)
+    prefix = output_dir / f"frame_{roi_packet.source_frame_id:06d}_roi_{roi_packet.roi_id:06d}"
+    cv2.imwrite(str(prefix) + "_raw.png", roi_packet.roi_image)
+    cv2.imwrite(str(prefix) + "_code.png", engine_result.code_roi)
+    cv2.imwrite(str(prefix) + "_processed.png", engine_result.processed)
+
+
+def detect_yolov2(
+    net,
+    frame: np.ndarray,
+    classes: list[str],
+    min_confidence: float,
+    nms_threshold: float,
+    input_size: int,
+    use_nms: bool,
+    swap_rb: bool,
+) -> tuple[list[Detection], dict[str, float | tuple[int, ...]]]:
+    t0 = time.perf_counter()
+    h, w = frame.shape[:2]
+
+    blob = cv2.dnn.blobFromImage(
+        frame,
+        scalefactor=1.0 / 255.0,
+        size=(input_size, input_size),
+        mean=(0, 0, 0),
+        swapRB=swap_rb,
+        crop=False,
+    )
+    t_blob = time.perf_counter()
+
+    net.setInput(blob)
+    predictions = net.forward()
+    t_forward = time.perf_counter()
+
+    predictions = np.array(predictions)
+    if predictions.ndim == 4:
+        predictions = predictions.reshape(predictions.shape[1], predictions.shape[-1])
+    elif predictions.ndim == 3:
+        predictions = predictions[0]
+
+    boxes: list[list[int]] = []
+    confidences: list[float] = []
+    class_ids: list[int] = []
+
+    if predictions.ndim == 2 and predictions.shape[1] > 5:
+        for i in range(predictions.shape[0]):
+            prob_arr = predictions[i][5:]
+            if prob_arr.size == 0:
+                continue
+
+            class_index = int(prob_arr.argmax(axis=0))
+            if class_index >= len(classes):
+                continue
+
+            confidence = float(prob_arr[class_index])
+            if confidence <= min_confidence:
+                continue
+
+            x_center = float(predictions[i][0]) * w
+            y_center = float(predictions[i][1]) * h
+            width_box = float(predictions[i][2]) * w
+            height_box = float(predictions[i][3]) * h
+
+            x1 = int(x_center - width_box * 0.5)
+            y1 = int(y_center - height_box * 0.5)
+            x2 = int(x_center + width_box * 0.5)
+            y2 = int(y_center + height_box * 0.5)
+            x1, y1, x2, y2 = clip_box(x1, y1, x2, y2, w, h)
+
+            bw = max(0, x2 - x1)
+            bh = max(0, y2 - y1)
+            if bw == 0 or bh == 0:
+                continue
+
+            boxes.append([x1, y1, bw, bh])
+            confidences.append(confidence)
+            class_ids.append(class_index)
+
+    detections: list[Detection] = []
+    if boxes:
+        if use_nms:
+            indices = cv2.dnn.NMSBoxes(boxes, confidences, min_confidence, nms_threshold)
+            selected_indices = np.array(indices).flatten() if len(indices) else []
+        else:
+            selected_indices = range(len(boxes))
+
+        for idx in selected_indices:
+            x, y, bw, bh = boxes[int(idx)]
+            class_id = class_ids[int(idx)]
+            detections.append(Detection(
+                class_id=class_id,
+                class_name=classes[class_id],
+                confidence=confidences[int(idx)],
+                bbox=(x, y, x + bw, y + bh),
+            ))
+
+    t_parse = time.perf_counter()
+    info = {
+        "shape": tuple(predictions.shape),
+        "blob_ms": (t_blob - t0) * 1000.0,
+        "forward_ms": (t_forward - t_blob) * 1000.0,
+        "parse_ms": (t_parse - t_forward) * 1000.0,
+        "total_ms": (t_parse - t0) * 1000.0,
+        "raw_max": float(np.max(predictions)) if predictions.size else 0.0,
+        "class_max": float(np.max(predictions[:, 5:])) if predictions.ndim == 2 and predictions.shape[1] > 5 else 0.0,
+        "raw_predictions": predictions,
+    }
+    return detections, info
+
+
+def best_label_detections(
+    detections: list[Detection],
+    label_name: str,
+    max_boxes: int,
+) -> list[Detection]:
+    labels = [
+        d for d in detections
+        if d.class_name.strip().lower() == label_name.strip().lower()
+    ]
+    labels.sort(
+        key=lambda d: (
+            d.confidence,
+            (d.bbox[2] - d.bbox[0]) * (d.bbox[3] - d.bbox[1]),
+        ),
+        reverse=True,
+    )
+    return labels[:max_boxes]
+
+
+def draw_detection(frame: np.ndarray, det: Detection, label_class: str) -> None:
+    x1, y1, x2, y2 = det.bbox
+    class_lower = det.class_name.lower()
+    color = (255, 255, 255)
+    if class_lower == "gaylord_stimato":
+        color = (255, 0, 255)
+    elif class_lower == label_class.lower():
+        color = (0, 255, 255)
+    elif class_lower == "gaylord":
+        color = (0, 255, 0)
+
+    thickness = 4 if class_lower in ("gaylord", "gaylord_stimato") else 2
+    cv2.rectangle(frame, (x1, y1), (x2, y2), color, thickness)
+    text = f"{det.class_name} {det.confidence:.2f}"
+    cv2.putText(
+        frame,
+        text,
+        (x1, max(20, y1 - 8)),
+        cv2.FONT_HERSHEY_SIMPLEX,
+        0.6,
+        color,
+        2,
+        cv2.LINE_AA,
+    )
+
+
+def infer_gaylords_from_labels(
+    detections: list[Detection],
+    frame_w: int,
+    frame_h: int,
+    label_class: str,
+    width_factor: float,
+    height_factor: float,
+    y_shift: float,
+) -> list[Detection]:
+    has_gaylord = any(d.class_name.lower() == "gaylord" for d in detections)
+    if has_gaylord:
+        return []
+
+    inferred: list[Detection] = []
+    for det in detections:
+        if det.class_name.lower() != label_class.lower():
+            continue
+
+        x1, y1, x2, y2 = det.bbox
+        bw = x2 - x1
+        bh = y2 - y1
+        if bw <= 0 or bh <= 0:
+            continue
+
+        cx = (x1 + x2) / 2.0
+        cy = (y1 + y2) / 2.0 + bh * y_shift
+        gw = bw * width_factor
+        gh = bh * height_factor
+
+        gx1 = int(cx - gw / 2.0)
+        gy1 = int(cy - gh / 2.0)
+        gx2 = int(cx + gw / 2.0)
+        gy2 = int(cy + gh / 2.0)
+        gx1, gy1, gx2, gy2 = clip_box(gx1, gy1, gx2, gy2, frame_w, frame_h)
+
+        inferred.append(Detection(
+            class_id=-1,
+            class_name="gaylord_stimato",
+            confidence=det.confidence,
+            bbox=(gx1, gy1, gx2, gy2),
+        ))
+
+    return inferred
+
+
+def draw_ocr_results(
+    frame: np.ndarray,
+    ocr_results: list[OcrResult],
+    max_age_sec: float = 5.0,
+) -> None:
+    now = time.perf_counter()
+    for result in ocr_results:
+        if now - result.timestamp > max_age_sec:
+            continue
+        if not result.text:
+            continue
+
+        x1, y1, x2, y2 = result.bbox
+        cv2.putText(
+            frame,
+            f"NUM: {result.text}",
+            (x1, min(frame.shape[0] - 5, y2 + 24)),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.7,
+            (0, 255, 255),
+            2,
+            cv2.LINE_AA,
+        )
+
+
+def draw_status(frame: np.ndarray, stats_text: list[str]) -> None:
+    y = 25
+    for line in stats_text:
+        cv2.putText(
+            frame,
+            line,
+            (10, y),
+            cv2.FONT_HERSHEY_SIMPLEX,
+            0.6,
+            (0, 0, 255),
+            2,
+            cv2.LINE_AA,
+        )
+        y += 24
+
+
+def draw_ocr_debug(ocr_frame: np.ndarray, text: str, preview_width: int) -> np.ndarray:
+    if len(ocr_frame.shape) == 2:
+        display = cv2.cvtColor(ocr_frame, cv2.COLOR_GRAY2BGR)
+    else:
+        display = ocr_frame.copy()
+
+    display = resize_preview(display, preview_width)
+    canvas_h = display.shape[0] + 70
+    canvas_w = max(display.shape[1], 500)
+    canvas = np.full((canvas_h, canvas_w, 3), 255, dtype=np.uint8)
+    canvas[:display.shape[0], :display.shape[1]] = display
+
+    shown_text = text if text else "(nessun codice)"
+    cv2.putText(
+        canvas,
+        f"OCR: {shown_text}",
+        (10, display.shape[0] + 45),
+        cv2.FONT_HERSHEY_SIMPLEX,
+        1.0,
+        (0, 0, 255),
+        2,
+        cv2.LINE_AA,
+    )
+    return canvas
+
+
+def resize_for_quality(frame: np.ndarray, target_width: int) -> np.ndarray:
+    if target_width <= 0 or frame.shape[1] <= target_width:
+        return frame
+    scale = target_width / frame.shape[1]
+    height = max(1, int(frame.shape[0] * scale))
+    return cv2.resize(frame, (target_width, height), interpolation=cv2.INTER_AREA)
+
+
+def estimate_sharpness(frame: np.ndarray, metric: str, resize_width: int) -> float:
+    small = resize_for_quality(frame, resize_width)
+    gray = cv2.cvtColor(small, cv2.COLOR_BGR2GRAY)
+    if metric == "tenengrad":
+        gx = cv2.Sobel(gray, cv2.CV_32F, 1, 0, ksize=3)
+        gy = cv2.Sobel(gray, cv2.CV_32F, 0, 1, ksize=3)
+        return float(np.mean(gx * gx + gy * gy))
+
+    lap = cv2.Laplacian(gray, cv2.CV_64F)
+    return float(lap.var())
+
+
+def format_stats(
+    shared: SharedState,
+    frame_buffer: LatestBuffer,
+    roi_buffer: LatestBuffer,
+    start_time: float,
+) -> list[str]:
+    stats = shared.snapshot_stats()
+    elapsed = max(0.001, time.perf_counter() - start_time)
+    fb = frame_buffer.stats()
+    rb = roi_buffer.stats()
+
+    avg_yolo = stats.yolo_total_ms / max(1, stats.yolo_cycles)
+    avg_forward = stats.yolo_forward_ms / max(1, stats.yolo_cycles)
+    avg_ocr = stats.ocr_total_ms / max(1, stats.ocr_cycles)
+    avg_display = stats.display_total_ms / max(1, stats.display_frames)
+    avg_quality = stats.quality_total_ms / max(1, stats.quality_cycles)
+    avg_sharpness = stats.quality_score_total / max(1, stats.quality_cycles)
+
+    return [
+        f"cap_fps={stats.capture_frames / elapsed:.1f}",
+        f"disp_fps={stats.display_frames / elapsed:.1f}",
+        f"quality yolo_in/reject={stats.yolo_submitted_frames}/{stats.quality_rejected_frames} "
+        f"last={stats.last_quality_score:.1f} avg={avg_sharpness:.1f} ms={avg_quality:.2f}",
+        f"yolo_fps={stats.yolo_cycles / elapsed:.1f} avg={avg_yolo:.1f}ms fwd={avg_forward:.1f}ms",
+        f"ocr_fps={stats.ocr_cycles / elapsed:.1f} avg={avg_ocr:.1f}ms",
+        f"display_avg={avg_display:.1f}ms",
+        f"frames id cap/yolo/ocr={stats.last_capture_frame_id}/{stats.last_yolo_frame_id}/{stats.last_ocr_frame_id}",
+        f"last_det={stats.last_detection_count} {stats.last_detection_summary}",
+        f"frame_buf push/pop/drop/skip={fb.pushed}/{fb.popped}/{fb.dropped_on_put}/{fb.skipped_on_pop}",
+        f"roi_buf push/pop/drop/skip={rb.pushed}/{rb.popped}/{rb.dropped_on_put}/{rb.skipped_on_pop}",
+    ]
+
+
+def yolo_worker(
+    stop_event: threading.Event,
+    frame_buffer: LatestBuffer,
+    roi_buffer: LatestBuffer,
+    shared: SharedState,
+    net,
+    classes: list[str],
+    args,
+    roi_id_gen: IdGenerator,
+) -> None:
+    slot_last_ocr: dict[tuple[int, int], float] = {}
+    last_ocr_submit = -999999.0
+    label_class_lower = args.label_class.strip().lower()
+
+    log("YOLO worker avviato")
+    while not stop_event.is_set():
+        packet = frame_buffer.get_latest_blocking(stop_event)
+        if packet is None:
+            continue
+
+        local_frame = packet.frame.copy()
+        detections, info = detect_yolov2(
+            net=net,
+            frame=local_frame,
+            classes=classes,
+            min_confidence=args.min_confidence,
+            nms_threshold=args.nms_threshold,
+            input_size=args.input_size,
+            use_nms=args.use_nms,
+            swap_rb=args.swap_rb,
+        )
+
+        result = DetectionResult(
+            frame_id=packet.frame_id,
+            timestamp=time.perf_counter(),
+            detections=detections,
+            inference_ms=float(info["total_ms"]),
+            blob_ms=float(info["blob_ms"]),
+            forward_ms=float(info["forward_ms"]),
+            parse_ms=float(info["parse_ms"]),
+            source_width=packet.width,
+            source_height=packet.height,
+        )
+        shared.set_latest_detection(result)
+
+        if args.debug_yolo_output:
+            log(
+                f"YOLO frame={packet.frame_id} shape={info['shape']} "
+                f"raw_max={info['raw_max']:.4f} class_max={info['class_max']:.4f} "
+                f"det={len(detections)}"
+            )
+            if args.debug_yolo_top > 0:
+                raw_predictions = info["raw_predictions"]
+                if (
+                    isinstance(raw_predictions, np.ndarray)
+                    and raw_predictions.ndim == 2
+                    and raw_predictions.shape[1] > 5
+                ):
+                    scores = raw_predictions[:, 5:]
+                    row_best = scores.max(axis=1)
+                    top_indices = np.argsort(row_best)[-args.debug_yolo_top:][::-1]
+                    for idx in top_indices:
+                        cls_scores = scores[idx]
+                        cls_parts = " ".join(
+                            f"{classes[i]}={float(cls_scores[i]):.4f}"
+                            for i in range(min(len(classes), cls_scores.shape[0]))
+                        )
+                        log(
+                            f"  raw[{int(idx)}] obj={float(raw_predictions[idx, 4]):.4f} "
+                            f"{cls_parts} box={tuple(float(v) for v in raw_predictions[idx, :4])}"
+                        )
+
+        debug = local_frame.copy()
+        debug_detections = list(detections)
+        if args.infer_gaylord_from_label:
+            inferred_gaylords = infer_gaylords_from_labels(
+                detections,
+                packet.width,
+                packet.height,
+                args.label_class,
+                args.inferred_gaylord_width_factor,
+                args.inferred_gaylord_height_factor,
+                args.inferred_gaylord_y_shift,
+            )
+            debug_detections.extend(inferred_gaylords)
+            if args.debug_inferred_gaylord and inferred_gaylords:
+                log(
+                    f"gaylord stimati frame={packet.frame_id}: "
+                    f"{[det.bbox for det in inferred_gaylords]}"
+                )
+        for det in debug_detections:
+            draw_detection(debug, det, args.label_class)
+        shared.set_debug_yolo_frame(resize_preview(debug, args.preview_width))
+
+        label_dets = best_label_detections(
+            detections,
+            args.label_class,
+            args.max_roi_per_frame,
+        )
+        now = time.perf_counter()
+        for det in label_dets:
+            if now - last_ocr_submit < args.ocr_submit_min_interval:
+                continue
+            if roi_buffer.depth() >= args.ocr_max_pending:
+                continue
+
+            x1, y1, x2, y2 = det.bbox
+            bw = x2 - x1
+            bh = y2 - y1
+            if bw < args.min_label_width or bh < args.min_label_height:
+                continue
+
+            slot_key = quantized_slot_key(det.bbox, args.slot_size)
+            if now - slot_last_ocr.get(slot_key, -999999.0) < args.ocr_cooldown_sec:
+                continue
+
+            rx1, ry1, rx2, ry2 = expand_box(
+                x1, y1, x2, y2,
+                packet.width,
+                packet.height,
+                pad_ratio=args.ocr_pad_ratio,
+            )
+            roi = local_frame[ry1:ry2, rx1:rx2]
+            if roi.size == 0:
+                continue
+
+            roi_copy = roi.copy()
+            roi_buffer.put(RoiPacket(
+                roi_id=roi_id_gen.next(),
+                source_frame_id=packet.frame_id,
+                timestamp=now,
+                class_name=label_class_lower,
+                confidence=det.confidence,
+                bbox=det.bbox,
+                roi_image=roi_copy,
+                width=roi_copy.shape[1],
+                height=roi_copy.shape[0],
+            ))
+            slot_last_ocr[slot_key] = now
+            last_ocr_submit = now
+
+    log("YOLO worker terminato")
+
+
+def ocr_worker(
+    stop_event: threading.Event,
+    roi_buffer: LatestBuffer,
+    shared: SharedState,
+    args,
+) -> None:
+    try:
+        engine = create_ocr_engine(args)
+    except Exception as exc:
+        log(f"OCR worker disabilitato: impossibile inizializzare {args.ocr_backend}: {exc}")
+        return
+
+    debug_dir = Path(args.save_ocr_roi_dir) if args.save_ocr_roi_dir else None
+
+    log(f"OCR worker avviato con backend {args.ocr_backend}")
+    try:
+        while not stop_event.is_set():
+            roi_packet = roi_buffer.get_latest_blocking(stop_event)
+            if roi_packet is None:
+                continue
+
+            t0 = time.perf_counter()
+            engine_result = engine.read_digits(roi_packet.roi_image)
+            ocr_ms = (time.perf_counter() - t0) * 1000.0
+            digits = engine_result.digits
+            raw_text = engine_result.raw_text
+            processed = engine_result.processed
+
+            if debug_dir is not None:
+                save_ocr_debug_images(debug_dir, roi_packet, engine_result)
+
+            if args.print_all_ocr:
+                log(
+                    f"OCR frame={roi_packet.source_frame_id} "
+                    f"raw='{raw_text.strip()}' digits='{digits}' ms={ocr_ms:.1f}"
+                )
+
+            result = OcrResult(
+                roi_id=roi_packet.roi_id,
+                source_frame_id=roi_packet.source_frame_id,
+                timestamp=time.perf_counter(),
+                text=digits if len(digits) >= args.ocr_min_digits else "",
+                raw_text=raw_text,
+                bbox=roi_packet.bbox,
+                ocr_ms=ocr_ms,
+            )
+            shared.add_ocr_result(result)
+            shared.set_debug_ocr_frame(processed, result.text or raw_text.strip())
+
+            if result.text:
+                log(
+                    f"Etichetta letta frame={result.source_frame_id} "
+                    f"roi={result.roi_id}: {result.text} ({ocr_ms:.1f} ms)"
+                )
+    finally:
+        close = getattr(engine, "close", None)
+        if close is not None:
+            close()
+
+    log("OCR worker terminato")
+
+
+def main() -> int:
+    args = parse_args()
+    if args.easyocr_worker:
+        return run_easyocr_worker(args)
+
+    require_file(args.weights, "File pesi YOLOv2")
+    require_file(args.config, "File config YOLOv2")
+    require_file(args.labels, "File labels")
+    classes = load_classes(args.labels)
+
+    cv2.setNumThreads(args.opencv_threads)
+    log(f"OpenCV version: {cv2.__version__}")
+    log(f"Classi: {classes}")
+    log(f"Backend richiesto: {args.backend}")
+
+    net = cv2.dnn.readNetFromDarknet(args.config, args.weights)
+    configure_net_backend(net, args.backend)
+
+    cap, source_name = open_capture(args.video)
+    if not cap.isOpened():
+        log("ERRORE: impossibile aprire la sorgente video")
+        return 1
+
+    frame_buffer = LatestBuffer(args.frame_buffer_size, "frames")
+    roi_buffer = LatestBuffer(args.roi_buffer_size, "roi")
+    shared = SharedState(args.ocr_history_size)
+    stop_event = threading.Event()
+    roi_id_gen = IdGenerator()
+    start_time = time.perf_counter()
+
+    yolo_thread = threading.Thread(
+        target=yolo_worker,
+        name="yolo-worker",
+        args=(stop_event, frame_buffer, roi_buffer, shared, net, classes, args, roi_id_gen),
+        daemon=True,
+    )
+    yolo_thread.start()
+    ocr_thread = None
+    if not args.no_ocr:
+        ocr_thread = threading.Thread(
+            target=ocr_worker,
+            name="ocr-worker",
+            args=(stop_event, roi_buffer, shared, args),
+            daemon=True,
+        )
+        ocr_thread.start()
+    else:
+        log("OCR disabilitato da --no-ocr")
+
+    if not args.no_display:
+        cv2.namedWindow("flywms capture", cv2.WINDOW_NORMAL)
+        if args.debug_yolo_window:
+            cv2.namedWindow("flywms yolo", cv2.WINDOW_NORMAL)
+        if args.debug_ocr_window or not args.no_ocr:
+            cv2.namedWindow("flywms ocr", cv2.WINDOW_NORMAL)
+        if args.quality_filter and args.debug_rejected_window:
+            cv2.namedWindow("flywms scartati", cv2.WINDOW_NORMAL)
+
+    frame_id = 0
+    last_stats_log = time.perf_counter()
+    video_fps = cap.get(cv2.CAP_PROP_FPS) or 0.0
+    frame_period = 1.0 / video_fps if args.realtime_playback and video_fps > 0 else 0.0
+
+    try:
+        while not stop_event.is_set():
+            loop_start = time.perf_counter()
+            t_read0 = time.perf_counter()
+            grabbed, frame = cap.read()
+            read_ms = (time.perf_counter() - t_read0) * 1000.0
+
+            if not grabbed or frame is None:
+                log("Fine stream o impossibile leggere il frame")
+                break
+
+            frame_id += 1
+            if args.max_frames > 0 and frame_id > args.max_frames:
+                log(f"Raggiunto --max-frames={args.max_frames}")
+                break
+
+            height, width = frame.shape[:2]
+            sharpness = 0.0
+            quality_ms = 0.0
+            quality_passed = True
+            if args.quality_filter or args.debug_quality_log:
+                t_quality0 = time.perf_counter()
+                sharpness = estimate_sharpness(
+                    frame,
+                    args.blur_metric,
+                    args.blur_resize_width,
+                )
+                quality_ms = (time.perf_counter() - t_quality0) * 1000.0
+                quality_passed = sharpness >= args.min_sharpness
+            if args.debug_quality_log:
+                log(
+                    f"quality frame={frame_id} sharpness={sharpness:.1f} "
+                    f"passed={quality_passed} ms={quality_ms:.2f}"
+                )
+
+            packet = FramePacket(
+                frame_id=frame_id,
+                timestamp=time.perf_counter(),
+                frame=frame,
+                width=width,
+                height=height,
+                source=source_name,
+            )
+            shared.add_capture_read(frame_id, read_ms)
+            if quality_passed:
+                frame_buffer.put(packet)
+            shared.add_quality_result(
+                score=sharpness,
+                passed=quality_passed,
+                elapsed_ms=quality_ms,
+                submitted_to_yolo=quality_passed,
+            )
+
+            t_display0 = time.perf_counter()
+            if not args.no_display:
+                display = frame.copy()
+                if args.quality_filter:
+                    quality_color = (0, 180, 0) if quality_passed else (0, 0, 255)
+                    cv2.putText(
+                        display,
+                        f"sharp={sharpness:.1f} {'OK' if quality_passed else 'BLUR'}",
+                        (20, 38),
+                        cv2.FONT_HERSHEY_SIMPLEX,
+                        1.0,
+                        quality_color,
+                        2,
+                        cv2.LINE_AA,
+                    )
+                latest_detection = shared.get_latest_detection()
+                if latest_detection is not None:
+                    display_detections = list(latest_detection.detections)
+                    if args.infer_gaylord_from_label:
+                        inferred_gaylords = infer_gaylords_from_labels(
+                            latest_detection.detections,
+                            latest_detection.source_width,
+                            latest_detection.source_height,
+                            args.label_class,
+                            args.inferred_gaylord_width_factor,
+                            args.inferred_gaylord_height_factor,
+                            args.inferred_gaylord_y_shift,
+                        )
+                        display_detections.extend(inferred_gaylords)
+                    for det in display_detections:
+                        draw_detection(display, det, args.label_class)
+
+                draw_ocr_results(display, shared.get_recent_ocr_results())
+                draw_status(display, format_stats(shared, frame_buffer, roi_buffer, start_time)[:5])
+                display = resize_preview(display, args.preview_width)
+                cv2.imshow("flywms capture", display)
+
+                debug_yolo, debug_ocr, debug_ocr_text = shared.get_debug_frames()
+                if args.debug_yolo_window and debug_yolo is not None:
+                    cv2.imshow("flywms yolo", debug_yolo)
+                if (args.debug_ocr_window or not args.no_ocr) and debug_ocr is not None:
+                    cv2.imshow(
+                        "flywms ocr",
+                        draw_ocr_debug(debug_ocr, debug_ocr_text, args.preview_width),
+                    )
+                if (
+                    args.quality_filter
+                    and args.debug_rejected_window
+                    and not quality_passed
+                ):
+                    rejected = frame.copy()
+                    cv2.putText(
+                        rejected,
+                        f"SCARTATO sharp={sharpness:.1f} < {args.min_sharpness:.1f}",
+                        (20, 38),
+                        cv2.FONT_HERSHEY_SIMPLEX,
+                        1.0,
+                        (0, 0, 255),
+                        2,
+                        cv2.LINE_AA,
+                    )
+                    rejected = resize_preview(rejected, args.preview_width)
+                    cv2.imshow("flywms scartati", rejected)
+
+                key = cv2.waitKey(1) & 0xFF
+                if key == ord("q"):
+                    log("Premuto q, uscita")
+                    break
+
+            display_ms = (time.perf_counter() - t_display0) * 1000.0
+            shared.add_display(display_ms)
+
+            now = time.perf_counter()
+            if now - last_stats_log >= args.stats_interval:
+                for line in format_stats(shared, frame_buffer, roi_buffer, start_time):
+                    log(line)
+                last_stats_log = now
+
+            if frame_period > 0:
+                elapsed = time.perf_counter() - loop_start
+                if elapsed < frame_period:
+                    time.sleep(frame_period - elapsed)
+
+    except KeyboardInterrupt:
+        log("Interrotto da tastiera")
+    finally:
+        if args.drain_seconds > 0:
+            log(f"Drain worker per {args.drain_seconds:.1f}s")
+            time.sleep(args.drain_seconds)
+        stop_event.set()
+        frame_buffer.wake_all()
+        roi_buffer.wake_all()
+        yolo_thread.join(timeout=3.0)
+        if ocr_thread is not None:
+            ocr_thread.join(timeout=3.0)
+        cap.release()
+        if not args.no_display:
+            cv2.destroyAllWindows()
+
+    log("=== riepilogo finale ===")
+    for line in format_stats(shared, frame_buffer, roi_buffer, start_time):
+        log(line)
+    return 0
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())
diff --git a/flywms_pipeline_spec.rtf b/flywms_pipeline_spec.rtf
new file mode 100644
index 0000000..fcb1b08
--- /dev/null
+++ b/flywms_pipeline_spec.rtf
@@ -0,0 +1,421 @@
+{\rtf1\ansi\deff0
+{\fonttbl{\f0 Calibri;}{\f1 Consolas;}}
+\fs24
+\b Specifica pipeline realtime flywms\par\b0
+\par
+\b Obiettivo\par\b0
+Costruire una pipeline realtime modulare per video/camera, con tre blocchi principali:\par
+\par
+{\f1 Capture/Preview\par
+YOLO Detection\par
+OCR\par}
+\par
+I blocchi lavorano in parallelo, ognuno al proprio ritmo. Se un blocco e' lento, non deve bloccare gli altri. I frame/ROI vecchi possono essere scartati per mantenere bassa la latenza.\par
+\par
+\b Principio chiave\par\b0
+Non elaboriamo necessariamente tutti i frame. Elaboriamo il piu' recente disponibile quando un worker e' libero.\par
+\par
+{\f1 meglio saltare frame vecchi\par
+che accumulare ritardo\par}
+\par
+\b Strutture dati\par\b0
+{\f1 FramePacket\par
+  frame_id: intero crescente\par
+  timestamp: tempo acquisizione\par
+  frame: immagine OpenCV BGR\par
+  width: larghezza\par
+  height: altezza\par
+  source: video/camera\par
+\par
+Detection\par
+  class_id\par
+  class_name\par
+  confidence\par
+  bbox: x1, y1, x2, y2\par
+\par
+DetectionResult\par
+  frame_id\par
+  timestamp\par
+  detections: lista Detection\par
+  inference_ms\par
+  source_width\par
+  source_height\par
+\par
+RoiPacket\par
+  roi_id\par
+  source_frame_id\par
+  timestamp\par
+  class_name\par
+  confidence\par
+  bbox\par
+  roi_image: crop BGR copiato\par
+  width\par
+  height\par
+\par
+OcrResult\par
+  roi_id\par
+  source_frame_id\par
+  timestamp\par
+  text\par
+  raw_text\par
+  confidence opzionale\par
+  bbox\par
+  ocr_ms\par}
+\par
+\b Buffer\par\b0
+Usiamo buffer thread-safe "latest", non FIFO classici.\par
+\par
+{\f1 LatestBuffer\par
+  max_size\par
+  put(item)\par
+  get_latest_blocking(stop_event)\par
+  get_latest_nowait()\par
+  stats()\par}
+\par
+Comportamento:\par
+\par
+{\f1 put(item):\par
+  se pieno, elimina elementi vecchi\par
+  inserisce item nuovo\par
+  notifica i worker in attesa\par
+\par
+get_latest_blocking():\par
+  se vuoto, aspetta\par
+  quando c'e' almeno un item:\par
+    prende il piu' recente\par
+    scarta gli elementi piu' vecchi\par
+    restituisce item\par}
+\par
+Buffer previsti:\par
+\par
+{\f1 frame_buffer: FramePacket\par
+roi_buffer: RoiPacket\par}
+\par
+Risultati condivisi:\par
+\par
+{\f1 SharedState\par
+  latest_detection_result\par
+  latest_ocr_results\par
+  counters/statistiche\par}
+\par
+SharedState va protetto con lock breve.\par
+\par
+\b Thread\par\b0
+{\f1 Main/Capture Thread\par
+  legge frame\par
+  genera FramePacket\par
+  mette FramePacket in frame_buffer\par
+  visualizza preview\par
+  opzionalmente disegna ultimi risultati disponibili\par
+\par
+YOLO Worker Thread\par
+  aspetta ultimo FramePacket\par
+  copia il frame fuori lock\par
+  esegue detection\par
+  salva DetectionResult in SharedState\par
+  crea RoiPacket per le etichette utili\par
+  mette RoiPacket in roi_buffer\par
+  aggiorna statistiche\par
+\par
+OCR Worker Thread\par
+  aspetta ultimo RoiPacket\par
+  analizza ROI\par
+  salva OcrResult in SharedState/log\par
+  aggiorna statistiche\par}
+\par
+\b Thread safety\par\b0
+Il lock deve proteggere solo operazioni piccole:\par
+\par
+{\f1 inserire riferimento nel buffer\par
+leggere riferimento dal buffer\par
+aggiornare contatori\par
+aggiornare latest result\par
+notify/wait\par}
+\par
+Il lock non deve coprire:\par
+\par
+{\f1 cap.read()\par
+frame.copy()\par
+YOLO forward\par
+OCR\par
+cv2.imshow\par
+conversioni immagine\par}
+\par
+\b Politica copie\par\b0
+Prima versione, robusta:\par
+\par
+{\f1 Capture:\par
+  non copia ogni frame\par
+  mette il frame letto nel FramePacket\par
+\par
+YOLO:\par
+  prende FramePacket\par
+  fa frame.copy() solo del frame che analizzera'\par
+\par
+OCR:\par
+  riceve ROI gia' copiata\par}
+\par
+Questo evita corruzione dati e mantiene copie limitate.\par
+\par
+\b Politica realtime\par\b0
+Se YOLO e' lento:\par
+\par
+{\f1 non analizza frame arretrati\par
+prende il piu' recente disponibile\par}
+\par
+Se OCR e' lento:\par
+\par
+{\f1 non legge ROI vecchie\par
+prende la ROI piu' recente disponibile\par}
+\par
+Questo riduce la latenza.\par
+\par
+\b Algoritmo pseudocodice - main/capture\par\b0
+{\f1 inizializza config\par
+carica labels\par
+carica modello YOLO\par
+apri video/camera\par
+\par
+crea stop_event\par
+crea frame_buffer max_size=N\par
+crea roi_buffer max_size=M\par
+crea shared_state\par
+\par
+avvia thread yolo_worker\par
+avvia thread ocr_worker\par
+\par
+frame_id = 0\par
+\par
+loop main:\par
+  grabbed, frame = cap.read()\par
+\par
+  se non grabbed:\par
+    stop_event.set()\par
+    break\par
+\par
+  frame_id += 1\par
+\par
+  packet = FramePacket(\par
+    frame_id=frame_id,\par
+    timestamp=now(),\par
+    frame=frame,\par
+    width=frame.width,\par
+    height=frame.height,\par
+    source=source_name\par
+  )\par
+\par
+  frame_buffer.put(packet)\par
+\par
+  detection_result = shared_state.get_latest_detection()\par
+  ocr_results = shared_state.get_recent_ocr_results()\par
+\par
+  display_frame = frame.copy()\par
+  draw detections compatibili su display_frame\par
+  draw OCR texts compatibili su display_frame\par
+\par
+  cv2.imshow("Capture", display_frame)\par
+\par
+  se premuto q:\par
+    stop_event.set()\par
+    break\par
+\par
+attendi chiusura worker\par
+rilascia video\par
+chiudi finestre\par}
+\par
+\b Algoritmo pseudocodice - YOLO worker\par\b0
+{\f1 function yolo_worker():\par
+  while not stop_event:\par
+    packet = frame_buffer.get_latest_blocking(stop_event)\par
+\par
+    se packet e' None:\par
+      continue\par
+\par
+    local_frame = packet.frame.copy()\par
+\par
+    t0 = now()\par
+    detections = run_yolo(local_frame)\par
+    inference_ms = now() - t0\par
+\par
+    result = DetectionResult(\par
+      frame_id=packet.frame_id,\par
+      timestamp=now(),\par
+      detections=detections,\par
+      inference_ms=inference_ms,\par
+      source_width=packet.width,\par
+      source_height=packet.height\par
+    )\par
+\par
+    shared_state.set_latest_detection(result)\par
+\par
+    per ogni detection in detections:\par
+      se detection.class_name non e' "etichetta":\par
+        continue\par
+\par
+      se bbox troppo piccola:\par
+        continue\par
+\par
+      se OCR cooldown per questa zona non e' scaduto:\par
+        continue\par
+\par
+      roi = crop local_frame usando bbox espansa\par
+      roi_copy = roi.copy()\par
+\par
+      roi_packet = RoiPacket(\par
+        roi_id=next_roi_id(),\par
+        source_frame_id=packet.frame_id,\par
+        timestamp=now(),\par
+        class_name=detection.class_name,\par
+        confidence=detection.confidence,\par
+        bbox=detection.bbox,\par
+        roi_image=roi_copy,\par
+        width=roi_copy.width,\par
+        height=roi_copy.height\par
+      )\par
+\par
+      roi_buffer.put(roi_packet)\par
+\par
+    aggiorna statistiche yolo\par}
+\par
+\b Algoritmo pseudocodice - OCR worker\par\b0
+{\f1 function ocr_worker():\par
+  while not stop_event:\par
+    roi_packet = roi_buffer.get_latest_blocking(stop_event)\par
+\par
+    se roi_packet e' None:\par
+      continue\par
+\par
+    t0 = now()\par
+    processed = preprocess_for_ocr(roi_packet.roi_image)\par
+    text, raw_text = run_ocr(processed)\par
+    ocr_ms = now() - t0\par
+\par
+    result = OcrResult(\par
+      roi_id=roi_packet.roi_id,\par
+      source_frame_id=roi_packet.source_frame_id,\par
+      timestamp=now(),\par
+      text=text,\par
+      raw_text=raw_text,\par
+      bbox=roi_packet.bbox,\par
+      ocr_ms=ocr_ms\par
+    )\par
+\par
+    shared_state.add_ocr_result(result)\par
+\par
+    se text valido:\par
+      logga risultato\par
+\par
+    aggiorna statistiche ocr\par}
+\par
+\b LatestBuffer pseudocodice\par\b0
+{\f1 class LatestBuffer:\par
+  items = deque(maxlen=max_size)\par
+  condition = Condition()\par
+  pushed = 0\par
+  popped = 0\par
+  dropped = 0\par
+\par
+  put(item):\par
+    with condition:\par
+      before = len(items)\par
+\par
+      se before == max_size:\par
+        dropped += 1\par
+\par
+      items.append(item)\par
+      pushed += 1\par
+      condition.notify_all()\par
+\par
+  get_latest_blocking(stop_event):\par
+    with condition:\par
+      while items e' vuoto and not stop_event.is_set():\par
+        condition.wait(timeout=0.1)\par
+\par
+      se stop_event.is_set():\par
+        return None\par
+\par
+      latest = items[-1]\par
+      skipped = len(items) - 1\par
+      dropped += skipped\par
+      items.clear()\par
+      popped += 1\par
+\par
+      return latest\par}
+\par
+\b SharedState pseudocodice\par\b0
+{\f1 class SharedState:\par
+  lock\par
+  latest_detection_result\par
+  ocr_results_deque maxlen=100\par
+  stats\par
+\par
+  set_latest_detection(result):\par
+    with lock:\par
+      latest_detection_result = result\par
+\par
+  get_latest_detection():\par
+    with lock:\par
+      return latest_detection_result\par
+\par
+  add_ocr_result(result):\par
+    with lock:\par
+      ocr_results_deque.append(result)\par
+\par
+  get_recent_ocr_results():\par
+    with lock:\par
+      return copia lista risultati recenti\par}
+\par
+\b Statistiche minime\par\b0
+Da stampare ogni N secondi:\par
+\par
+{\f1 capture_fps\par
+yolo_fps\par
+ocr_fps\par
+frame_buffer pushed/popped/dropped\par
+roi_buffer pushed/popped/dropped\par
+avg_yolo_ms\par
+avg_ocr_ms\par
+last_frame_id\par
+last_yolo_frame_id\par
+last_ocr_frame_id\par
+latency_yolo_ms\par
+latency_ocr_ms\par}
+\par
+\b Configurazione iniziale\par\b0
+{\f1 --video testhd.mp4\par
+--frame-buffer-size 10\par
+--roi-buffer-size 20\par
+--min-confidence 0.30\par
+--label-class etichetta\par
+--max-roi-per-frame 2\par
+--ocr-min-digits 2\par
+--preview-width 1280\par
+--backend cpu|cuda|cuda-fp16\par
+--debug-yolo-window\par
+--debug-ocr-window\par}
+\par
+\b Prima versione: cosa evitare\par\b0
+Per la prima versione eviterei:\par
+\par
+{\f1 Dear PyGui\par
+PaddleOCR\par
+modello YOLO moderno\par
+tracking complesso\par}
+\par
+Prima validiamo la pipeline con OpenCV window + YOLO attuale + OCR attuale. Poi sostituiamo i motori uno alla volta.\par
+\par
+\b Decisione architetturale\par\b0
+Questa pipeline privilegia:\par
+\par
+{\f1 bassa latenza\par
+fluidita' realtime\par
+uso controllato della memoria\par
+misurabilita'\par}
+\par
+e sacrifica:\par
+\par
+{\f1 analisi completa di ogni frame\par
+ordine perfetto di output per ogni frame\par}
+\par
+Per un sistema operativo realtime e' il compromesso giusto.\par
+}
diff --git a/pip-before-opencv-cuda.txt b/pip-before-opencv-cuda.txt
new file mode 100644
index 0000000..9fce375
--- /dev/null
+++ b/pip-before-opencv-cuda.txt
@@ -0,0 +1 @@
+imutils==0.5.4