Update preprocessing.py

preprocessing.py

@@ -15,7 +15,8 @@ class VideoPreprocessor:
     def preprocess_video(self, video_path):
         cap = cv2.VideoCapture(video_path)
         frames = []
-    
+
+        # Utilize mediapipe's GPU acceleration if available
         with self.mp_face_mesh.FaceMesh(
             static_image_mode=False,
             max_num_faces=1,
@@ -27,53 +28,72 @@ class VideoPreprocessor:
                 ret, frame = cap.read()
                 if not ret:
                     break
-    
+
+                # Convert the BGR image to RGB
                 rgb_frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+
+                # Process the frame and get the facial landmarks
                 results = face_mesh.process(rgb_frame)
-    
+
                 if results.multi_face_landmarks:
+                    # Get the landmarks for the first face
                     face_landmarks = results.multi_face_landmarks[0]
+
                     try:
+                        # Extract lip landmarks
                         lip_landmarks = [face_landmarks.landmark[i] for i in self.LIP_INDICES]
+
+                        # Extract bounding box around the lips
                         h, w, _ = frame.shape
                         x_coords = [int(landmark.x * w) for landmark in lip_landmarks]
                         y_coords = [int(landmark.y * h) for landmark in lip_landmarks]
-    
+
                         x_min, x_max = max(0, min(x_coords)), min(w, max(x_coords))
                         y_min, y_max = max(0, min(y_coords)), min(h, max(y_coords))
-    
+
                         if x_max > x_min and y_max > y_min:
+                            # Crop the lip region
                             lip_frame = frame[y_min:y_max, x_min:x_max]
+
+                            # Resize to 160x160
                             lip_frame_resized = cv2.resize(lip_frame, (85, 85))
-                            lip_frame_gray = tf.image.rgb_to_grayscale(tf.cast(lip_frame_resized, tf.float32))
+
+                            # Convert to grayscale using TensorFlow
+                            lip_frame_gray = tf.image.rgb_to_grayscale(lip_frame_resized)
+
                             frames.append(lip_frame_gray)
                     except Exception as e:
                         print(f"Error processing frame: {e}")
-                        continue
+                        continue  # Skip this frame
                 else:
                     print("No face landmarks detected in frame.")
-    
+
         cap.release()
-    
+
         if not frames:
             print("No frames extracted during preprocessing.")
-            return None
-    
+            return None  # Return None to indicate failure
+
+        # Stack frames into a tensor
         frames = tf.stack(frames)
-    
+
+        # Adjust frames to match expected input length
         desired_num_frames = 75
         num_frames = frames.shape[0]
         if num_frames < desired_num_frames:
+            # Pad frames with zeros
             padding = tf.zeros((desired_num_frames - num_frames, 85, 85, 1), dtype=tf.float32)
             frames = tf.concat([frames, padding], axis=0)
         elif num_frames > desired_num_frames:
+            # Truncate frames to desired_num_frames
             frames = frames[:desired_num_frames]
-    
-        mean = tf.math.reduce_mean(tf.cast(frames, tf.float32))
+
+        # Normalize the frames
+        mean = tf.math.reduce_mean(frames)
         std = tf.math.reduce_std(tf.cast(frames, tf.float32))
         normalized_frames = tf.cast((frames - mean), tf.float32) / std
-    
-        return normalized_frames
+
+        return normalized_frames  # Return TensorFlow tensor
         print(f"Type of lip_frame_gray: {lip_frame_gray.dtype}")
         print(f"Type of frames[0]: {frames[0].dtype}")  # Trước khi stack