fix: improve content detection in trim_whitespace

- Add thresholding to better separate content from background
- Invert image so content becomes white for getbbox detection
- Fix issue where white and transparent areas weren't being trimmed

pdf_cleaner.py

@@ -97,18 +97,26 @@ def trim_whitespace(image_path: str, is_final: bool = False) -> bool:
         # Open image with PIL
         image = Image.open(image_path)
         
-        # Create flattened version with white background for finding bounds
+        # If image has transparency, flatten it first
-        flattened = Image.new('RGB', image.size, 'white')
         if image.mode == 'RGBA':
-            flattened.paste(image, mask=image.split()[3])
+            # Create a white background
-        else:
+            background = Image.new('RGB', image.size, 'white')
-            flattened.paste(image)
+            # Paste using alpha channel as mask
+            background.paste(image, mask=image.split()[3])
+            image = background
             
-        # Convert to grayscale for finding bounds
+        # Convert to grayscale
-        flattened = flattened.convert('L')
+        image = image.convert('L')
         
-        # Get the bounding box of non-white pixels
+        # Threshold to make all light pixels white and everything else black
-        bbox = flattened.getbbox()
+        # This helps with finding content bounds
+        image = image.point(lambda x: 255 if x > 250 else 0)
+        
+        # Invert so content is white on black background
+        image = Image.eval(image, lambda x: 255 - x)
+        
+        # Get the bounding box of content (now white pixels)
+        bbox = image.getbbox()
         if not bbox:
             print(f"Warning: No content found in {image_path}")
             return False
@@ -122,8 +130,9 @@ def trim_whitespace(image_path: str, is_final: bool = False) -> bool:
         x2 = min(width, x2 + padding)
         y2 = min(height, y2 + padding)
         
-        # Crop the original image (preserving transparency)
+        # Open original image again and crop it using the bounds
-        cropped = image.crop((x1, y1, x2, y2))
+        original = Image.open(image_path)
+        cropped = original.crop((x1, y1, x2, y2))
         
         # Save the cropped image, optimizing only if this is the final operation
         cropped.save(image_path, "PNG", optimize=is_final)