Update app.py

app.py

@@ -160,31 +160,45 @@ def get_beard_mask_fast(pil_image: Image.Image, exclude_mask: np.ndarray):
             if int(cls) == 0:  # beard class
                 m = results[0].masks.data[i].cpu().numpy()
                 m = cv2.resize(m, (orig_w, orig_h), interpolation=cv2.INTER_LINEAR)
-                mask = np.maximum(mask, (m > 0.35).astype(np.float32))
+                mask = np.maximum(mask, (m > 0.32).astype(np.float32))   # slightly lower threshold
 
     # Remove protected areas
     mask = np.maximum(mask - exclude_mask, 0)
 
-    if mask.sum() > 30:
-        # ====================== IMPROVED ROUNDING + NO EXTRA SPACE ======================
+    if mask.sum() > 25:   # lowered a bit
+        # ====================== HIGHLY IMPROVED ROUNDING ======================
         
-        # Step 1: Light erosion to remove extra pixels
-        kernel_erode = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
-        mask = cv2.erode(mask, kernel_erode, iterations=1)
+        # 1. Aggressive erosion to kill extra space
+        kernel_erode = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (7, 7))
+        mask = cv2.erode(mask, kernel_erode, iterations=2)
 
-        # Step 2: Strong close with big elliptical kernel → Best rounding
-        kernel_close = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (11, 11))
-        mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel_close, iterations=2)
+        # 2. Strong close with very big kernel for perfect round corners
+        kernel_close = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (13, 13))
+        mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernel_close, iterations=3)
 
-        # Step 3: Light open to clean noise
+        # 3. Clean noise
         kernel_open = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
         mask = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernel_open, iterations=1)
 
-        # Step 4: Soft blur for natural edges
-        mask = cv2.GaussianBlur(mask, (7, 7), 1.8)
+        # 4. Contour based smoothing (Best for round & tight beard shape)
+        contours, _ = cv2.findContours((mask > 0.1).astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_NONE)
+        if contours:
+            smooth_mask = np.zeros_like(mask, dtype=np.float32)
+            for cnt in contours:
+                if cv2.contourArea(cnt) > 50:  # ignore very small noise
+                    epsilon = 0.008 * cv2.arcLength(cnt, True)   # lower = more round
+                    approx = cv2.approxPolyDP(cnt, epsilon, True)
+                    cv2.drawContours(smooth_mask, [approx], -1, 1.0, thickness=cv2.FILLED)
+            mask = smooth_mask
 
-        # Step 5: Tight threshold → No extra space
-        mask = (mask > 0.25).astype(np.float32)
+        # 5. Final soft blur
+        mask = cv2.GaussianBlur(mask, (9, 9), 2.0)
+
+        # 6. Very tight final threshold (no extra space)
+        mask = (mask > 0.28).astype(np.float32)
+
+        # 7. Final light erosion for extra safety
+        mask = cv2.erode(mask, cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3)), iterations=1)
 
     return mask
 
@@ -259,8 +273,8 @@ def process_face_whitening(input_image: Image.Image):
     hair_mask, exclude_mask, mustache_mask = get_hair_and_exclude_masks(img_resized)
     beard_mask = get_beard_mask_fast(img_resized, exclude_mask)
     
-    # Combine beard + mustache with slight reduction on mustache
-    beard_mask = np.maximum(beard_mask, mustache_mask * 0.90)
+    # Combine with mustache
+    beard_mask = np.maximum(beard_mask, mustache_mask * 0.88)
 
     final_resized = apply_strong_grey_hair(img_resized, hair_mask, beard_mask)
     final_img = final_resized.resize((ow, oh), Image.LANCZOS)