Update app.py

app.py

@@ -52,13 +52,21 @@ def get_embedding(face):
     Preprocess the face and compute embedding using ArcFace ONNX model.
     Fixed to match NHWC format (1, 112, 112, 3).
     """
+    # Ensure face is not empty
+    if face.size == 0 or face.shape[0] < 10 or face.shape[1] < 10:
+        return None
+        
     img = cv2.resize(face, (112, 112))
     img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
-    img = (img - 127.5) / 128.0
-    img = np.expand_dims(img.astype(np.float32), axis=0)  # (1, 112, 112, 3)
+    
+    # Normalize to [-1, 1] range for better feature extraction
+    img = (img.astype(np.float32) - 127.5) / 128.0
+    img = np.expand_dims(img, axis=0)  # (1, 112, 112, 3)
 
     emb = arcface_sess.run(None, {arcface_input: img})[0][0]
-    emb = emb / np.linalg.norm(emb)
+    
+    # L2 normalization
+    emb = emb / (np.linalg.norm(emb) + 1e-8)
     return emb
 
 # ------------------------------
@@ -127,10 +135,26 @@ def register_face(name, image):
         else:
             img = cv2.cvtColor(img, cv2.COLOR_RGB2BGR)
 
-        # Compute embedding
-        emb = get_embedding(img)
-        if emb is None or np.isnan(emb).any():
-            return f"❌ Failed to compute embedding for {name}. Try another image."
+        # Detect face first using YOLO to ensure good quality
+        results = yolo.predict(source=img, conf=0.25, verbose=False)
+        
+        face_found = False
+        for r in results:
+            boxes = r.boxes
+            if len(boxes) > 0:
+                # Use the first detected face
+                box = boxes[0]
+                x1, y1, x2, y2 = map(int, box.xyxy[0].tolist())
+                crop = img[y1:y2, x1:x2]
+                
+                # Compute embedding from cropped face
+                emb = get_embedding(crop)
+                if emb is not None and not np.isnan(emb).any():
+                    face_found = True
+                    break
+        
+        if not face_found:
+            return f"❌ No clear face detected for {name}. Please upload a clearer image with a visible face."
 
         # Add to known faces dictionary
         known_faces[name] = emb
@@ -211,7 +235,9 @@ def detect_and_recognize(image):
                     score = cosine_similarity([emb], [ref_emb])[0][0]
                     if score > best_score:
                         best_name, best_score = name, score
-                if best_score < 0.45:
+                
+                # Lower threshold for better matching (0.35 instead of 0.45)
+                if best_score < 0.35:
                     best_name = "Unknown"
                 cv2.rectangle(img, (x1, y1), (x2, y2), (0,255,0), 2)
                 cv2.putText(img, f"{best_name} ({best_score:.2f})", (x1, y1-10),