Update ocr_engine.py

ocr_engine.py

@@ -9,24 +9,37 @@ def extract_weight_from_image(pil_img):
     try:
         img = np.array(pil_img)
 
-        # Step 1: Convert to grayscale
+        # Convert to grayscale
         gray = cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)
 
-        # Step 2: Apply adaptive threshold to handle lighting
-        thresh = cv2.adaptiveThreshold(gray, 255, cv2.ADAPTIVE_THRESH_GAUSSIAN_C,
-                                       cv2.THRESH_BINARY_INV, 11, 2)
+        # Apply bilateral filter to reduce noise while keeping edges
+        filtered = cv2.bilateralFilter(gray, 11, 17, 17)
 
-        # Step 3: Dilate to make digits thicker
-        kernel = np.ones((2, 2), np.uint8)
-        dilated = cv2.dilate(thresh, kernel, iterations=1)
+        # Apply binary threshold
+        _, thresh = cv2.threshold(filtered, 150, 255, cv2.THRESH_BINARY_INV)
 
-        # Step 4: OCR on the preprocessed image
-        result = reader.readtext(dilated, detail=0)
-        text = " ".join(result).strip()
-        print("OCR Text:", text)
+        # Find contours
+        contours, _ = cv2.findContours(thresh, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
 
-        # Step 5: Match numeric values like 52.30 or 003.25
-        match = re.search(r"\b\d{2,4}\.?\d{0,2}\b", text)
+        if not contours:
+            return "No weight detected", 0.0
+
+        # Get the largest contour assuming it's the display area
+        largest_contour = max(contours, key=cv2.contourArea)
+        x, y, w, h = cv2.boundingRect(largest_contour)
+
+        # Add padding
+        pad = 10
+        x, y = max(x - pad, 0), max(y - pad, 0)
+        cropped = gray[y:y+h+pad, x:x+w+pad]
+
+        # OCR on cropped area
+        result = reader.readtext(cropped, detail=0)
+        combined = " ".join(result)
+        print("Detected Text:", combined)
+
+        # Match weight patterns like 52.30 or 003.25
+        match = re.search(r"\b\d{2,4}\.?\d{0,2}\b", combined)
         if match:
             return match.group(), 95.0
         else: