Update app.py

app.py

@@ -112,52 +112,62 @@ def filter_nested_boxes(boxes, containment_thresh=0.9):
 def split_double_lines(crop_img, logs):
     """
     Analyzes a crop to see if it accidentally contains TWO lines of text.
-    Uses Horizontal Projection Profile.
-    Returns: List of crops (either [original] or [top_half, bottom_half])
+    Includes 'Descender Protection' to prevent cutting off tails (y, g, p).
     """
     # 1. Binarize
     gray = cv2.cvtColor(crop_img, cv2.COLOR_RGB2GRAY)
-    # Otsu's thresholding for dynamic contrast
     _, thresh = cv2.threshold(gray, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
 
-    # 2. Horizontal Projection (Sum of white pixels per row)
+    # 2. Horizontal Projection
     h_proj = np.sum(thresh, axis=1)
     
-    # 3. Normalize projection
+    # Normalize
     max_val = np.max(h_proj)
-    if max_val == 0: return [crop_img] # Empty image
+    if max_val == 0: return [crop_img]
     h_proj = h_proj / max_val
     
-    # 4. Find Peaks (Lines of text) and Valleys (Space between lines)
-    # We look for peaks with a certain prominence
-    peaks, _ = find_peaks(h_proj, height=0.2, distance=15)
+    # 3. Find Peaks
+    # Increased distance to 25 to prevent finding peaks within the same letter height
+    peaks, _ = find_peaks(h_proj, height=0.2, distance=25)
     
     if len(peaks) < 2:
-        return [crop_img] # Likely just one line
+        return [crop_img]
         
-    # If we have 2+ clear peaks, we check the "valley" between them
-    # Simple logic: Find the deepest point between the two major peaks
-    if len(peaks) >= 2:
-        # Get the first two major peaks
-        p1, p2 = peaks[0], peaks[1]
-        
-        # Look at the region between peaks
-        valley_region = h_proj[p1:p2]
-        if len(valley_region) == 0: return [crop_img]
-        
-        min_val = np.min(valley_region)
-        min_idx = np.argmin(valley_region) + p1
-        
-        # STRICT CHECK: Only split if the valley is truly empty (or noise)
-        # If the valley still has > 30% ink density of the peak, it might just be a messy 'y' or 'g'
-        if min_val < 0.3: 
-            logs.append(f"   -> ✂️ Refinement: Split double line at Y={min_idx}")
-            top_crop = crop_img[0:min_idx, :]
-            bot_crop = crop_img[min_idx:, :]
-            return [top_crop, bot_crop]
+    # 4. Analyze the Valley
+    p1, p2 = peaks[0], peaks[1]
+    valley_region = h_proj[p1:p2]
+    
+    if len(valley_region) == 0: return [crop_img]
+    
+    min_val = np.min(valley_region)
+    min_idx = np.argmin(valley_region) + p1
+    
+    # ==========================
+    # 🛡️ SAFETY GUARDRAILS
+    # ==========================
+    
+    # CHECK 1: Valley Depth
+    # Handwriting lines often touch. If the valley isn't DEEP (very low ink), don't split.
+    # We lowered the threshold to 0.15 (15% ink density)
+    if min_val > 0.15:
+        return [crop_img]
 
-    return [crop_img]
+    # CHECK 2: Edge Protection (The "Descender" Check)
+    # If the split point is too close to the bottom (e.g., > 75% down the image),
+    # it's almost certainly chopping off tails (y, g, p, q), not separating a new line.
+    total_height = crop_img.shape[0]
+    split_ratio = min_idx / total_height
+    
+    if split_ratio < 0.20 or split_ratio > 0.75:
+        logs.append(f"   -> ⚠️ Refinement: Prevented split at {int(split_ratio*100)}% (Likely descenders)")
+        return [crop_img]
 
+    # If we pass checks, perform the split
+    logs.append(f"   -> ✂️ Refinement: Split double line at Y={min_idx}")
+    top_crop = crop_img[0:min_idx, :]
+    bot_crop = crop_img[min_idx:, :]
+    
+    return [top_crop, bot_crop]
 # ==========================================
 # ⛓️ PIPELINE STEP: MERGING & ORDERING
 # ==========================================