Adding Cascade Solution

app.py

@@ -75,20 +75,79 @@ def find_and_straighten_document(image):
     return perspective_transform(image, box)
 
 def correct_orientation(image):
+    """
+    Corrects the orientation of an image using a robust cascade method.
+    1. Tries fast OSD (Orientation and Script Detection).
+    2. If OSD fails, falls back to analyzing word bounding boxes.
+    """
+    print("--- Running Orientation Check ---")
+    
+    # --- METHOD 1: Fast Orientation and Script Detection (OSD) ---
     try:
-        osd = pytesseract.image_to_osd(image, output_type=pytesseract.Output.DICT)
+        osd = pytesseract.image_to_osd(image, output_type=pytesseract.Output.DICT, timeout=5)
         rotation = osd['rotate']
-        if rotation in [90, 180, 270]:
+        print(f"OSD check successful. Detected rotation: {rotation} degrees.")
+        if rotation > 0:
+            # Tesseract's rotation is counter-clockwise
             if rotation == 90:
-                rotated_image = cv2.rotate(image, cv2.ROTATE_90_COUNTERCLOCKWISE)
+                return cv2.rotate(image, cv2.ROTATE_90_COUNTERCLOCKWISE)
             elif rotation == 180:
-                rotated_image = cv2.rotate(image, cv2.ROTATE_180)
+                return cv2.rotate(image, cv2.ROTATE_180)
             else: # 270
-                rotated_image = cv2.rotate(image, cv2.ROTATE_90_CLOCKWISE)
-            return rotated_image
+                return cv2.rotate(image, cv2.ROTATE_90_CLOCKWISE)
+        return image # Return as is if rotation is 0
     except Exception as e:
-        st.warning(f"OSD check failed: {e}. Returning original image.")
-    return image
+        print(f"OSD check failed: {e}. Falling back to word box analysis.")
+
+    # --- METHOD 2: Fallback using Word Bounding Box Analysis ---
+    # This method is slower but more robust for images with little text.
+    best_rotation = 0
+    max_horizontal_boxes = -1
+    
+    # Pre-process image once for all rotations
+    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
+    thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)[1]
+
+    orientations = {
+        0: thresh,
+        90: cv2.rotate(thresh, cv2.ROTATE_90_CLOCKWISE),
+        180: cv2.rotate(thresh, cv2.ROTATE_180),
+        270: cv2.rotate(thresh, cv2.ROTATE_90_COUNTERCLOCKWISE)
+    }
+
+    for angle, rotated_img in orientations.items():
+        try:
+            data = pytesseract.image_to_data(rotated_img, output_type=pytesseract.Output.DICT, timeout=5)
+            horizontal_boxes = 0
+            num_boxes = len(data['level'])
+            for i in range(num_boxes):
+                # We only consider word-level boxes (level 5) with some confidence
+                if data['level'][i] == 5 and int(data['conf'][i]) > 10: 
+                    w = data['width'][i]
+                    h = data['height'][i]
+                    if w > h: # Check if the box is horizontal
+                        horizontal_boxes += 1
+            
+            print(f"  Rotation {angle}°: Found {horizontal_boxes} horizontal word boxes.")
+            
+            if horizontal_boxes > max_horizontal_boxes:
+                max_horizontal_boxes = horizontal_boxes
+                best_rotation = angle
+        except Exception as e:
+            print(f"  Word box analysis failed for rotation {angle}°: {e}")
+            continue
+
+    print(f"--> Best rotation found at {best_rotation} degrees.")
+
+    # Apply the best rotation to the ORIGINAL color image
+    if best_rotation == 90:
+        return cv2.rotate(image, cv2.ROTATE_90_CLOCKWISE)
+    elif best_rotation == 180:
+        return cv2.rotate(image, cv2.ROTATE_180)
+    elif best_rotation == 270:
+        return cv2.rotate(image, cv2.ROTATE_90_COUNTERCLOCKWISE)
+    else: # 0 degrees
+        return image
 
 # ==============================================================================
 # NEW AND IMPROVED: Table Structure Recognition using OpenCV for Drawing