Split

manga-translator.py

@@ -238,8 +238,31 @@ def split_panel_box(image_bgr, bbox_xyxy, bubble_quads=None):
     peak_x_relative = np.argmax(vertical_projection[search_start:search_end]) + search_start
     peak_val = vertical_projection[peak_x_relative]
     
-    # Detect panel border with more sensitive threshold (0.40 instead of 0.60)
-    # This catches boxes with vertical lines even if they're not super dark
+    # Find ALL significant peaks, not just the first
+    # This helps detect internal gaps between text regions (left vs right bubbles)
+    threshold_val = h * 255 * 0.25  # Very sensitive threshold
+    significant_peaks = []
+    
+    for x_rel in range(search_start, search_end):
+        if vertical_projection[x_rel] > threshold_val:
+            significant_peaks.append((x_rel, vertical_projection[x_rel]))
+    
+    # If we have multiple peaks, find the largest gap between text regions
+    # by looking for the valley (lowest projection value) between peaks
+    if len(significant_peaks) > 1:
+        # Find the minimum value (gap) between the first and last peak
+        min_proj_val = np.min(vertical_projection[search_start:search_end])
+        min_proj_idx = np.argmin(vertical_projection[search_start:search_end]) + search_start
+        
+        # Use the valley point as the split, not the peak
+        # This more accurately separates left-aligned vs right-aligned content
+        if min_proj_val < threshold_val * 0.6:  # Valley is clearly a gap
+            split_x_absolute = x1 + min_proj_idx
+            box_left = (x1, y1, split_x_absolute, y2)
+            box_right = (split_x_absolute, y1, x2, y2)
+            return box_left, box_right, split_x_absolute
+    
+    # Fallback: if the main peak is significant enough, use it
     if peak_val > (h * 255 * 0.40):
         split_x_absolute = x1 + peak_x_relative
         box_left = (x1, y1, split_x_absolute, y2)
@@ -289,14 +312,14 @@ def split_bubble_if_multiple_columns(indices, ocr, bid=None, use_aggressive_thre
     
     if use_aggressive_thresholds:
         # Relaxed thresholds for fallback splitting after failed panel border
-        threshold1 = 10.0   # Very low absolute threshold for fallback
-        threshold2 = med_h * 0.8  # Even lower relative threshold
-        min_gap = 5.0
+        threshold1 = 80.0   # Increased from 10 to require significant gaps
+        threshold2 = med_h * 1.2  # Increased from 0.8
+        min_gap = 40.0  # Increased from 5
     else:
-        # Normal thresholds
-        threshold1 = 50.0   # Absolute threshold: 50 pixels
-        threshold2 = med_h * 2.0  # Relative threshold: 2x line height
-        min_gap = 25.0
+        # Normal thresholds - very conservative to avoid breaking valid bubbles
+        threshold1 = 120.0  # Increased from 50 - require very large gaps
+        threshold2 = med_h * 3.0  # Increased from 2.0 - require 3x line height
+        min_gap = 60.0  # Increased from 25
     
 
     
@@ -304,6 +327,12 @@ def split_bubble_if_multiple_columns(indices, ocr, bid=None, use_aggressive_thre
         split_idx = max_gap_idx
         left_indices = [item[0] for item in sorted_items[:split_idx]]
         right_indices = [item[0] for item in sorted_items[split_idx:]]
+        
+        # Additional safety: don't split if one side only has 1 detection
+        # This prevents breaking up valid bubbles with just a few words
+        if len(left_indices) < 2 or len(right_indices) < 2:
+            return None
+        
         return left_indices, right_indices
     
     return None
@@ -637,6 +666,10 @@ def group_tokens(ocr, image_shape, gap_px=18, bbox_padding=3):
     hs = [max(1.0, b[3] - b[1]) for b in boxes]
     med_h = float(np.median(hs)) if hs else 12.0
     dist_thresh = max(20.0, med_h * 2.2)
+    
+    # Adaptive vertical gap: allow up to 1.0x median line height for vertical gaps
+    # This handles cases where lines are spaced further apart (e.g., multi-line bubbles)
+    adaptive_gap_y = max(gap_px, med_h * 1.0)
 
     p = list(range(n))
 
@@ -651,7 +684,13 @@ def group_tokens(ocr, image_shape, gap_px=18, bbox_padding=3):
 
     for i in range(n):
         for j in range(i + 1, n):
-            if overlap_or_near(boxes[i], boxes[j], gap=gap_px):
+            # Use adaptive gap for vertical spacing, fixed gap for horizontal
+            ax1, ay1, ax2, ay2 = boxes[i]
+            bx1, by1, bx2, by2 = boxes[j]
+            gap_x = max(0, max(ax1, bx1) - min(ax2, bx2))
+            gap_y = max(0, max(ay1, by1) - min(ay2, by2))
+            
+            if gap_x <= gap_px and gap_y <= adaptive_gap_y:
                 unite(i, j)
                 continue
             cx1, cy1 = centers[i]
@@ -681,8 +720,12 @@ def group_tokens(ocr, image_shape, gap_px=18, bbox_padding=3):
             continue
 
         x1, y1, x2, y2 = ub
-        x1 = max(0, x1 - bbox_padding); y1 = max(0, y1 - bbox_padding)
-        x2 = min(iw - 1, x2 + bbox_padding); y2 = min(ih - 1, y2 + bbox_padding)
+        
+        # Adaptive padding: scale with median line height to ensure all boundary chars are captured
+        # Use max of fixed padding or line-height-based padding
+        adaptive_pad = max(bbox_padding, int(round(med_h * 0.35)))
+        x1 = max(0, x1 - adaptive_pad); y1 = max(0, y1 - adaptive_pad)
+        x2 = min(iw - 1, x2 + adaptive_pad); y2 = min(ih - 1, y2 + adaptive_pad)
 
         bubbles[bid] = lines
         bubble_boxes[bid] = (x1, y1, x2, y2)