Sync from GitHub: bbd3a8da74f29857d203fdb7913c964597eb127b

inference.py

@@ -69,7 +69,7 @@ class InferenceProcessor:
     """Handles VLM inference, validation, and result processing"""
     
     @staticmethod
-    def preprocess_image(image_path: str, enhance: bool = None) -> Image.Image:
+    def preprocess_image(image_path: str, enhance: bool = None) -> Tuple[Image.Image, str]:
         """Load, enhance (optional), and resize image if needed
         
         Args:
@@ -77,7 +77,7 @@ class InferenceProcessor:
             enhance: Whether to enhance image quality before processing (None=use config default)
         
         Returns:
-            Preprocessed PIL Image ready for VLM inference
+            Tuple of (PIL Image ready for VLM, path to image file for YOLO)
         """
         # Use config default if not specified
         if enhance is None:
@@ -85,7 +85,6 @@ class InferenceProcessor:
         
         # Step 1: Enhance image if enabled
         enhanced_path = image_path
-        cleanup_enhanced = False
         
         if enhance:
             try:
@@ -95,7 +94,6 @@ class InferenceProcessor:
                     scale=ENHANCEMENT_SCALE,
                     model_name=ENHANCEMENT_MODEL
                 )
-                cleanup_enhanced = (enhanced_path != image_path)
             except Exception as e:
                 print(f"⚠️ Enhancement failed: {str(e)}, using original image")
                 enhanced_path = image_path
@@ -103,13 +101,6 @@ class InferenceProcessor:
         # Step 2: Load image
         image = Image.open(enhanced_path).convert("RGB")
         
-        # Cleanup enhanced temp file if created
-        if cleanup_enhanced:
-            try:
-                os.unlink(enhanced_path)
-            except:
-                pass
-        
         # Step 3: Resize if too large
         if max(image.size) > MAX_IMAGE_SIZE:
             ratio = MAX_IMAGE_SIZE / max(image.size)
@@ -117,7 +108,8 @@ class InferenceProcessor:
             image = image.resize(new_size, Image.LANCZOS)
             print(f"🔄 Image resized to {new_size}")
         
-        return image
+        # Return both PIL Image and path (path will be cleaned up by caller)
+        return image, enhanced_path
     
     @staticmethod
     def run_vlm_extraction(image: Image.Image) -> Tuple[str, float]:
@@ -348,12 +340,12 @@ class InferenceProcessor:
         
         # Step 1: Preprocess image (with optional enhancement)
         t1 = time.time()
-        image = InferenceProcessor.preprocess_image(image_path, enhance=enhance)
+        image, enhanced_image_path = InferenceProcessor.preprocess_image(image_path, enhance=enhance)
         timing_breakdown['image_preprocessing'] = round(time.time() - t1, 3)
         
-        # Step 2: YOLO Detection
+        # Step 2: YOLO Detection (use enhanced image path for consistency)
         t2 = time.time()
-        signature_info, stamp_info, signature_conf, stamp_conf = model_manager.detect_sign_stamp(image_path)
+        signature_info, stamp_info, signature_conf, stamp_conf = model_manager.detect_sign_stamp(enhanced_image_path)
         timing_breakdown['yolo_detection'] = round(time.time() - t2, 3)
         
         # Step 3: VLM Extraction
@@ -361,10 +353,17 @@ class InferenceProcessor:
         vlm_output, vlm_latency = InferenceProcessor.run_vlm_extraction(image)
         timing_breakdown['vlm_inference'] = round(vlm_latency, 3)
         
-        # Clean up image
+        # Clean up image and enhanced file if it was created
         image.close()
         del image
         
+        # Cleanup enhanced temp file if created
+        if enhanced_image_path != image_path:
+            try:
+                os.unlink(enhanced_image_path)
+            except:
+                pass
+        
         # Step 4: Parse JSON
         t4 = time.time()
         raw_json = InferenceProcessor.extract_json_from_output(vlm_output)