Update emotion.py

emotion.py

@@ -1,39 +1,77 @@
 import time
 import threading
 from dataclasses import dataclass, field
-from typing import Optional, List, Tuple, Dict
+from typing import Optional, Dict, Tuple, List
 
 import gradio as gr
 import numpy as np
-from PIL import Image
-
-import cv2
-import mediapipe as mp
-from mediapipe.tasks import python
-from mediapipe.tasks.python import vision
+from PIL import Image, ImageDraw
 
 from deepface import DeepFace
 
 
+# ===== 避免 css 未定義造成 NameError =====
+css = ""
+
+
 # =========================
-# 參數（可微調）
+# 參數（優先順：快）
 # =========================
-ANALYZE_EVERY_SEC = 1.5          # ⭐ 建議 1.5~2.0：HF CPU 比較順
+ANALYZE_EVERY_SEC = 1.2     # ✅ 降頻率（原 0.8）
+DOWNSAMPLE_W = 480          # ✅ 降解析（原 720）
 EMA_ALPHA = 0.45
-CONF_TH = 40.0                   # DeepFace emotion confidence 門檻
-SWITCH_CONFIRM_SEC = 2.5         # 連續穩定多久才算完成
-ZOOM_FACTOR = 1.5                # ✅ 保留：只用在「顯示」
-UI_TARGET_W = 720                # ✅ 回傳畫面縮小寬度，提升順暢度
+CONF_TH = 40.0
+SWITCH_CONFIRM_SEC = 2.5
+IOU_TH = 0.25
+ZOOM_FACTOR = 1.25          # ✅ 稍微降 zoom（減少失真與負擔）
+
+DETECTOR_BACKEND = "opencv" # ✅ 改快（原 mtcnn）
+ALIGN_FACE = True
 
-# MediaPipe face detector 門檻（太高很容易漏）
-MP_MIN_DET_CONF = 0.3
+# 防誤判（保留）
+MIN_FACE_AREA_RATIO = 0.05
+MIN_DET_CONF = 0.90
+
+
+# =========================
+# Warm-up（優先順：按 Record 不要等）
+# =========================
+_WARMED = False
+
+def _warmup_models():
+    global _WARMED
+    if _WARMED:
+        return
+    _WARMED = True
+    try:
+        dummy = np.zeros((224, 224, 3), dtype=np.uint8)
+        # detector warmup
+        try:
+            DeepFace.extract_faces(
+                img_path=dummy,
+                detector_backend=DETECTOR_BACKEND,
+                enforce_detection=False,
+                align=ALIGN_FACE,
+            )
+        except Exception:
+            pass
+        # emotion model warmup
+        try:
+            DeepFace.analyze(
+                img_path=dummy,
+                actions=["emotion"],
+                enforce_detection=False,
+            )
+        except Exception:
+            pass
+    except Exception:
+        pass
 
-# 模型路徑（確認 repo 中模型真的在這裡）
-MP_MODEL_PATH = "blaze_face_short_range.tflite"
+threading.Thread(target=_warmup_models, daemon=True).start()
 
 
 # =========================
-# 情緒中文 + 分數映射（7 -> 5）
+# 情緒中文 + 分數映射
 # =========================
 EMO_ZH = {
     "angry": "生氣",
@@ -72,26 +110,26 @@ SCORE_CHOICES = [SCORE_LABELS[i] for i in [1, 2, 3, 4, 5]]
 # =========================
 @dataclass
 class AppState:
-    mode: str = "攝影機辨識"
-
     running: bool = True
     finished: bool = False
     last_analyze_t: float = 0.0
 
-    # smoothing
+    track_bbox: Optional[Tuple[int, int, int, int]] = None
+
     ema_scores: Optional[np.ndarray] = None
     emo_keys: List[str] = field(default_factory=list)
 
-    # stable/candidate
     stable_emo: str = "unknown"
     candidate_emo: Optional[str] = None
     candidate_since: Optional[float] = None
 
-    # final
     final_emo: Optional[str] = None
     final_conf: Optional[float] = None
     final_score: int = 3
 
+    last_face_t: float = 0.0
+    face_in_box: bool = False
+
 
 # =========================
 # UI HTML
@@ -109,86 +147,46 @@ def _camera_hint_before_start() -> str:
         "接著，請按「錄製（Record）」開始。"
     )
 
-def _camera_hint_need_face() -> str:
-    return _hint_html(
-        "⚠️ 目前未偵測到人臉，請將臉部移至畫面中央、靠近鏡頭，並避免瀏海遮住眉眼，才可進行情緒辨識。"
-    )
-
 def _camera_hint_done() -> str:
     return _hint_html("已完成辨識，可按「重新攝影機辨識」再次進行辨識。")
 
 def _camera_hint_stopped() -> str:
     return _hint_html("已停止辨識。可按「重新攝影機辨識」再試一次。")
 
-def _score_radio_value(score: int) -> str:
-    score = int(score)
-    score = min(5, max(1, score))
-    return SCORE_LABELS[score]
-
 def _result_card_html(emo_key: str, conf: Optional[float]) -> str:
     zh = EMO_ZH.get(emo_key, emo_key)
     score = EMO_TO_SCORE.get(emo_key, 3)
     conf_txt = "" if conf is None else f"{conf:.1f}%"
     score_desc = SCORE_LABELS[score].split("（")[1].rstrip("）")
-
     return f"""
-    <div style="
-        border-radius:16px;
-        padding:20px;
-        border:1px solid #E6E6E6;
-        background:#FFFFFF;
-        color:#000000;
-    ">
-      <div style="font-size:14px;color:#000000;margin-bottom:8px;">
-        辨識結果
-      </div>
-
-      <div style="
-        font-size:42px;
-        font-weight:800;
-        line-height:1.1;
-        margin-bottom:14px;
-        color:#000000;
-      ">
-        {zh}
-      </div>
-
-      <div style="font-size:16px;color:#000000;margin-bottom:6px;">
-        心情���數：{score}（{score_desc}）
-      </div>
-
-      <div style="font-size:14px;color:#000000;">
-        信心值：{conf_txt}
-      </div>
+    <div style="border-radius:16px;padding:20px;border:1px solid #E6E6E6;background:#FFFFFF;color:#000000;">
+      <div style="font-size:14px;color:#000000;margin-bottom:8px;">辨識結果</div>
+      <div style="font-size:42px;font-weight:800;line-height:1.1;margin-bottom:14px;color:#000000;">{zh}</div>
+      <div style="font-size:16px;color:#000000;margin-bottom:6px;">心情分數：{score}（{score_desc}）</div>
+      <div style="font-size:14px;color:#000000;">信心值：{conf_txt}</div>
     </div>
     """
 
+def _score_radio_value(score: int) -> str:
+    score = int(score)
+    score = min(5, max(1, score))
+    return SCORE_LABELS[score]
+
 
 # =========================
-# 影像處理（保命 + 顯示優化）
+# 影像工具
 # =========================
-def _to_rgb_uint8(frame: np.ndarray) -> np.ndarray:
-    """統一成 MediaPipe Tasks 最穩格式：RGB、uint8、0~255、contiguous。"""
+def _downsample_rgb(frame: np.ndarray, target_w: int) -> np.ndarray:
     if frame is None:
+        return None
+    h, w = frame.shape[:2]
+    if w <= target_w:
         return frame
-
-    # RGBA -> RGB
-    if frame.ndim == 3 and frame.shape[2] == 4:
-        frame = frame[:, :, :3]
-
-    # 灰階 -> RGB
-    if frame.ndim == 2:
-        frame = np.stack([frame, frame, frame], axis=-1)
-
-    # dtype / range 修正
-    if frame.dtype != np.uint8:
-        mx = float(np.max(frame)) if frame.size else 0.0
-        if mx <= 1.5:  # 多半是 0~1
-            frame = (frame * 255.0).clip(0, 255).astype(np.uint8)
-        else:          # 多半是 0~255 但 float
-            frame = np.clip(frame, 0, 255).astype(np.uint8)
-
-    return np.ascontiguousarray(frame)
+    scale = target_w / w
+    new_h = max(1, int(h * scale))
+    img = Image.fromarray(frame)
+    img = img.resize((target_w, new_h))
+    return np.array(img)
 
 def _zoom_center(frame: np.ndarray, zoom: float) -> np.ndarray:
     if frame is None or zoom <= 1.0:
@@ -198,199 +196,166 @@ def _zoom_center(frame: np.ndarray, zoom: float) -> np.ndarray:
     new_h = int(h / zoom)
     start_x = (w - new_w) // 2
     start_y = (h - new_h) // 2
-    cropped = frame[start_y : start_y + new_h, start_x : start_x + new_w]
-    return np.ascontiguousarray(cropped)
-
-def _resize_for_ui(frame_rgb: np.ndarray, target_w: int = UI_TARGET_W) -> np.ndarray:
-    """回傳給 Gradio 的畫面縮到固定寬度，減少傳輸/渲染負擔。"""
-    if frame_rgb is None:
-        return frame_rgb
-    h, w = frame_rgb.shape[:2]
-    if w <= target_w:
-        return frame_rgb
-    scale = target_w / w
-    new_h = max(1, int(h * scale))
-    return cv2.resize(frame_rgb, (target_w, new_h), interpolation=cv2.INTER_AREA)
+    return frame[start_y:start_y + new_h, start_x:start_x + new_w]
 
 
 # =========================
-# MediaPipe Tasks Face Detector
+# 引導框（永遠畫）
 # =========================
-BaseOptions = python.BaseOptions
-FaceDetector = vision.FaceDetector
-FaceDetectorOptions = vision.FaceDetectorOptions
-VisionRunningMode = vision.RunningMode
-
-_face_detector = FaceDetector.create_from_options(
-    FaceDetectorOptions(
-        base_options=BaseOptions(model_asset_path=MP_MODEL_PATH),
-        running_mode=VisionRunningMode.IMAGE,
-        min_detection_confidence=MP_MIN_DET_CONF,
-    )
-)
+GUIDE_W_RATIO = 0.55
+GUIDE_H_RATIO = 0.70
+GUIDE_Y_OFFSET = 0.00
 
-def mp_tasks_detect_and_draw(frame_rgb: np.ndarray):
-    """
-    - 最大臉策略（bbox 面積最大）
-    - 有臉：畫白框 + 紅點 keypoints
-    - 回 bbox_px (x, y, w, h) 供 ROI 裁切
-    """
-    frame_rgb = _to_rgb_uint8(frame_rgb)
-    h, w = frame_rgb.shape[:2]
-
-    mp_image = mp.Image(image_format=mp.ImageFormat.SRGB, data=frame_rgb)
-    result = _face_detector.detect(mp_image)
-
-    if not result.detections:
-        return frame_rgb, None
-
-    best_det = None
-    best_area = -1.0
-    best_bbox = None
-
-    # 最大臉：bbox 面積最大
-    for det in result.detections:
-        bbox = det.bounding_box
-        area = float(bbox.width) * float(bbox.height)
-        if area > best_area:
-            best_area = area
-            best_det = det
-            best_bbox = bbox
-
-    if best_det is None or best_bbox is None:
-        return frame_rgb, None
-
-    lx = int(best_bbox.origin_x)
-    ly = int(best_bbox.origin_y)
-    bw = int(best_bbox.width)
-    bh = int(best_bbox.height)
-
-    # clamp
-    lx = max(0, min(w - 1, lx))
-    ly = max(0, min(h - 1, ly))
-    bw = max(1, min(w - lx, bw))
-    bh = max(1, min(h - ly, bh))
-
-    bgr = cv2.cvtColor(frame_rgb, cv2.COLOR_RGB2BGR)
+def _guide_box(frame: np.ndarray) -> Tuple[int, int, int, int]:
+    h, w = frame.shape[:2]
+    gw = int(w * GUIDE_W_RATIO)
+    gh = int(h * GUIDE_H_RATIO)
+    cx = w // 2
+    cy = int(h * (0.5 + GUIDE_Y_OFFSET))
+    x1 = max(0, cx - gw // 2)
+    y1 = max(0, cy - gh // 2)
+    x2 = min(w - 1, cx + gw // 2)
+    y2 = min(h - 1, cy + gh // 2)
+    return x1, y1, x2, y2
+
+def _bbox_center_in_box(bbox_xywh: Tuple[float, float, float, float], box_xyxy: Tuple[int, int, int, int]) -> bool:
+    x, y, w, h = bbox_xywh
+    cx = x + w / 2.0
+    cy = y + h / 2.0
+    x1, y1, x2, y2 = box_xyxy
+    return (x1 <= cx <= x2) and (y1 <= cy <= y2)
+
+def _draw_guide_overlay(frame_rgb: np.ndarray, detected_in_box: bool) -> np.ndarray:
+    img = Image.fromarray(frame_rgb)
+    draw = ImageDraw.Draw(img)
+
+    x1, y1, x2, y2 = _guide_box(frame_rgb)
+    color = (0, 200, 0) if detected_in_box else (220, 0, 0)
 
-    # 白框
-    cv2.rectangle(bgr, (lx, ly), (lx + bw, ly + bh), (255, 255, 255), 3)
+    h, w = frame_rgb.shape[:2]
+    thickness = max(3, int(min(w, h) * 0.008))
+    for t in range(thickness):
+        draw.rectangle([x1 - t, y1 - t, x2 + t, y2 + t], outline=color)
 
-    # 紅點：五官特徵點（相對座標）
-    for kp in best_det.keypoints:
-        cx = int(kp.x * w)
-        cy = int(kp.y * h)
-        cv2.circle(bgr, (cx, cy), 6, (0, 0, 255), -1)
+    msg = "已偵測到人臉，開始辨識情緒中…" if detected_in_box else "未偵測到人臉，請把臉放進框內"
+    tx = x1
+    ty = max(0, y1 - int(thickness * 6))
+    draw.rectangle([tx, ty, tx + 360, ty + 32], fill=(255, 255, 255))
+    draw.text((tx + 6, ty + 6), msg, fill=(0, 0, 0))
 
-    drawn = cv2.cvtColor(bgr, cv2.COLOR_BGR2RGB)
-    return drawn, (lx, ly, bw, bh)
+    return np.array(img)
 
 
 # =========================
-# 背景推論（避免卡頓）
+# Tracking / IoU
 # =========================
-_ANALYZE_LOCK = threading.Lock()
-_ANALYZE_THREAD: Optional[threading.Thread] = None
-_PENDING_FACE: Optional[np.ndarray] = None
-_LAST_ANALYZE: Dict[str, Optional[float]] = {"ts": 0.0, "conf": None, "emo": "unknown"}
-
-def _analyze_worker():
-    """背景 thread：取最新 face_roi 做 DeepFace.analyze，更新 _LAST_ANALYZE。"""
-    global _PENDING_FACE, _ANALYZE_THREAD, _LAST_ANALYZE
-
-    while True:
-        with _ANALYZE_LOCK:
-            face = _PENDING_FACE
-            _PENDING_FACE = None
-
-        if face is None:
-            break
-
-        try:
-            r = DeepFace.analyze(
-                img_path=face,
-                actions=["emotion"],
-                enforce_detection=False,
-                detector_backend="skip",  # ⭐ 關鍵：你已經裁好 ROI
-            )
-            if isinstance(r, list):
-                r = r[0]
-            emo_dict = r.get("emotion", None)
-
-            if isinstance(emo_dict, dict) and len(emo_dict) > 0:
-                top_key = max(emo_dict, key=lambda k: float(emo_dict[k]))
-                top_conf = float(emo_dict[top_key])
-                _LAST_ANALYZE = {"ts": time.time(), "emo": top_key, "conf": top_conf}
-        except Exception:
-            # 失敗就保持上一次結果
-            pass
-
-    _ANALYZE_THREAD = None
-
-
-def _enqueue_analyze(face_roi: np.ndarray):
-    """把最新 ROI 丟給背景 thread，若 thread 沒在跑就啟動。"""
-    global _PENDING_FACE, _ANALYZE_THREAD
-
-    with _ANALYZE_LOCK:
-        _PENDING_FACE = face_roi
-
-    if _ANALYZE_THREAD is None:
-        _ANALYZE_THREAD = threading.Thread(target=_analyze_worker, daemon=True)
-        _ANALYZE_THREAD.start()
+def _iou(a: Tuple[int, int, int, int], b: Tuple[int, int, int, int]) -> float:
+    ax, ay, aw, ah = a
+    bx, by, bw, bh = b
+    ax2, ay2 = ax + aw, ay + ah
+    bx2, by2 = bx + bw, by + bh
+
+    ix1, iy1 = max(ax, bx), max(ay, by)
+    ix2, iy2 = min(ax2, bx2), min(ay2, by2)
+    iw, ih = max(0, ix2 - ix1), max(0, iy2 - iy1)
+    inter = iw * ih
+    if inter <= 0:
+        return 0.0
+    union = aw * ah + bw * bh - inter
+    return inter / max(1e-6, union)
+
+def _pick_face_by_tracking(faces: List[Dict], prev_bbox: Optional[Tuple[int, int, int, int]]):
+    if not faces:
+        return None, None
+    cand = []
+    for f in faces:
+        fa = f.get("facial_area", {}) or {}
+        x = int(fa.get("x", 0) or 0)
+        y = int(fa.get("y", 0) or 0)
+        w = int(fa.get("w", 0) or 0)
+        h = int(fa.get("h", 0) or 0)
+        if w > 0 and h > 0 and f.get("face") is not None:
+            cand.append((f, (x, y, w, h), w * h))
+
+    if not cand:
+        return None, None
+
+    if prev_bbox is not None:
+        best = None
+        best_i = 0.0
+        for f, bb, area in cand:
+            i = _iou(prev_bbox, bb)
+            if i > best_i:
+                best_i = i
+                best = (f, bb)
+        if best is not None and best_i >= IOU_TH:
+            return best[0], best[1]
+
+    cand.sort(key=lambda x: x[2], reverse=True)
+    return cand[0][0], cand[0][1]
 
 
-# =========================
-# state helper
-# =========================
 def _reset_state_for_camera(st: AppState) -> AppState:
     st.running = True
     st.finished = False
     st.last_analyze_t = 0.0
-
+    st.track_bbox = None
     st.ema_scores = None
     st.emo_keys = []
-
     st.stable_emo = "unknown"
     st.candidate_emo = None
     st.candidate_since = None
-
     st.final_emo = None
     st.final_conf = None
     st.final_score = 3
+    st.last_face_t = 0.0
+    st.face_in_box = False
     return st
 
 
+def _is_real_face(pick: Dict, bbox_full_xywh: Tuple[float, float, float, float], frame_hw: Tuple[int, int]) -> bool:
+    fh, fw = frame_hw
+    x, y, w, h = bbox_full_xywh
+    area_ratio = (w * h) / max(1.0, float(fw * fh))
+    if area_ratio < MIN_FACE_AREA_RATIO:
+        return False
+    det_conf = pick.get("confidence", None)
+    if det_conf is not None:
+        try:
+            if float(det_conf) < MIN_DET_CONF:
+                return False
+        except Exception:
+            pass
+    return True
+
+
 # =========================
-# streaming callback
+# Stream（優先順）
 # =========================
 def on_stream(frame_rgb: np.ndarray, st: AppState):
-    # stopped / finished
     if st.finished or (not st.running):
         cam_visible = False
-
         if st.final_emo:
-            result_html = _result_card_html(st.final_emo, st.final_conf)
-            mood_val = _score_radio_value(EMO_TO_SCORE.get(st.final_emo, 3))
-            hint_html = _camera_hint_done()
-        else:
-            result_html = _hint_html("尚未完成辨識。")
-            mood_val = _score_radio_value(3)
-            hint_html = _camera_hint_stopped()
-
+            return (
+                gr.update(visible=cam_visible),
+                gr.update(value=_result_card_html(st.final_emo, st.final_conf)),
+                gr.update(value=_score_radio_value(EMO_TO_SCORE.get(st.final_emo, 3))),
+                gr.update(value=_camera_hint_done()),
+                st,
+                gr.update(value="重新攝影機辨識"),
+            )
         return (
-            gr.update(value=None, visible=cam_visible),
-            gr.update(value=result_html),
-            gr.update(value=mood_val),
-            gr.update(value=hint_html),
+            gr.update(visible=cam_visible),
+            gr.update(value=_hint_html("尚未完成辨識。")),
+            gr.update(value=_score_radio_value(3)),
+            gr.update(value=_camera_hint_stopped()),
             st,
             gr.update(value="重新攝影機辨識"),
         )
 
-    # no frame yet
     if frame_rgb is None:
         return (
-            gr.update(value=None, visible=True),
+            gr.update(visible=True),
             gr.update(value=_hint_html("等待影像…（尚未開始串流）")),
             gr.update(value=_score_radio_value(3)),
             gr.update(value=_camera_hint_before_start()),
@@ -398,93 +363,159 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
             gr.update(value="重新攝影機辨識"),
         )
 
-    # 統一格式
-    frame_rgb = _to_rgb_uint8(frame_rgb)
+    frame_rgb = _zoom_center(frame_rgb, ZOOM_FACTOR)
+    now = time.time()
+
+    # ✅ 先畫框（永遠立即有框），不等 detector
+    detected_recent = st.face_in_box and (now - st.last_face_t) < 1.0
+    annotated_fast = _draw_guide_overlay(frame_rgb, detected_recent)
+
+    # ✅ 如果還沒到分析時間，直接回傳畫框（超順）
+    if now - st.last_analyze_t < ANALYZE_EVERY_SEC:
+        return (
+            gr.update(visible=True, value=annotated_fast),
+            gr.update(value=_hint_html("已偵測到人臉，開始辨識情緒中…" if detected_recent else "未偵測到人臉，請把臉放進框內")),
+            gr.update(value=_score_radio_value(3)),
+            gr.update(value=_camera_hint_before_start()),
+            st,
+            gr.update(value="重新攝影機辨識"),
+        )
+
+    # 到時間才做一次重工作（低頻）
+    st.last_analyze_t = now
+    small = _downsample_rgb(frame_rgb, DOWNSAMPLE_W)
 
-    # ✅ 偵測用原圖（穩）
-    frame_drawn, bbox_px = mp_tasks_detect_and_draw(frame_rgb)
+    try:
+        faces = DeepFace.extract_faces(
+            img_path=small,
+            detector_backend=DETECTOR_BACKEND,
+            enforce_detection=True,
+            align=ALIGN_FACE,
+        )
+    except Exception:
+        faces = []
 
-    # ✅ 顯示用 zoom（你要保留 zoom 畫面）
-    frame_show = _zoom_center(frame_drawn if bbox_px is not None else frame_rgb, ZOOM_FACTOR)
-    frame_show = _resize_for_ui(frame_show, UI_TARGET_W)
+    pick, bbox = _pick_face_by_tracking(faces, st.track_bbox)
 
-    # 沒有人臉：不畫框/點 + 提示
-    if bbox_px is None:
+    if pick is None or bbox is None or pick.get("face") is None:
+        st.track_bbox = None
+        st.face_in_box = False
+        st.last_face_t = now
+        st.stable_emo = "unknown"
+        annotated = _draw_guide_overlay(frame_rgb, False)
         return (
-            gr.update(value=frame_show, visible=True),
-            gr.update(value=_hint_html("尚未偵測到人臉")),
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("未偵測到人臉，請把臉放進框內")),
             gr.update(value=_score_radio_value(3)),
-            gr.update(value=_camera_hint_need_face()),
+            gr.update(value=_camera_hint_before_start()),
             st,
             gr.update(value="重新攝影機辨識"),
         )
 
-    now = time.time()
+    sh, sw = small.shape[:2]
+    fh, fw = frame_rgb.shape[:2]
+    sx = fw / max(1, sw)
+    sy = fh / max(1, sh)
+    x, y, w, h = bbox
+    bbox_full = (x * sx, y * sy, w * sx, h * sy)
+
+    if not _is_real_face(pick, bbox_full, (fh, fw)):
+        st.track_bbox = None
+        st.face_in_box = False
+        st.last_face_t = now
+        annotated = _draw_guide_overlay(frame_rgb, False)
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("未偵測到人臉，請把臉放進框內")),
+            gr.update(value=_score_radio_value(3)),
+            gr.update(value=_camera_hint_before_start()),
+            st,
+            gr.update(value="重新攝影機辨識"),
+        )
+
+    guide = _guide_box(frame_rgb)
+    in_box = _bbox_center_in_box(bbox_full, guide)
+
+    st.last_face_t = now
+    st.face_in_box = in_box
+    st.track_bbox = bbox
 
-    # 到分析時間：只排隊給背景 thread（不在 on_stream 裡跑重推論）
-    if now - st.last_analyze_t >= ANALYZE_EVERY_SEC:
-        st.last_analyze_t = now
-
-        x, y, bw, bh = bbox_px
-        pad_x = int(bw * 0.10)
-        pad_y = int(bh * 0.10)
-        H, W = frame_rgb.shape[:2]
-        x1 = max(0, x - pad_x)
-        y1 = max(0, y - pad_y)
-        x2 = min(W, x + bw + pad_x)
-        y2 = min(H, y + bh + pad_y)
-        face_roi = frame_rgb[y1:y2, x1:x2].astype(np.uint8)
-
-        _enqueue_analyze(face_roi)
-
-    # 取「最後一次背景推論結果」
-    top_key = _LAST_ANALYZE.get("emo", "unknown") or "unknown"
-    top_conf = _LAST_ANALYZE.get("conf", None)
-
-    # 還沒有推論到結果：先用目前穩定值顯示
-    if top_conf is None:
-        show = st.stable_emo or "unknown"
+    if not in_box:
+        annotated = _draw_guide_overlay(frame_rgb, False)
         return (
-            gr.update(value=frame_show, visible=True),
-            gr.update(value=_result_card_html(show, st.final_conf)),
-            gr.update(value=_score_radio_value(EMO_TO_SCORE.get(show, 3))),
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("已偵測到人臉，但請把臉放進框內")),
+            gr.update(value=_score_radio_value(3)),
             gr.update(value=_camera_hint_before_start()),
             st,
             gr.update(value="重新攝影機辨識"),
         )
 
-    # 低信心就不切換，顯示目前穩定值
+    # 只有臉在框內才分析情緒
+    face = pick["face"]
+    if face.dtype != np.uint8:
+        face = np.clip(face * 255.0, 0, 255).astype(np.uint8)
+
+    try:
+        result = DeepFace.analyze(
+            img_path=face,
+            actions=["emotion"],
+            enforce_detection=True,
+        )
+    except Exception:
+        annotated = _draw_guide_overlay(frame_rgb, False)
+        st.face_in_box = False
+        st.last_face_t = now
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("未偵測到人臉，請把臉放進框內")),
+            gr.update(value=_score_radio_value(3)),
+            gr.update(value=_camera_hint_before_start()),
+            st,
+            gr.update(value="重新攝影機辨識"),
+        )
+
+    if isinstance(result, list):
+        result = result[0]
+    emo_dict = result.get("emotion", None)
+    if not isinstance(emo_dict, dict) or len(emo_dict) == 0:
+        annotated = _draw_guide_overlay(frame_rgb, True)
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("已偵測到人臉，開始辨識情緒中…")),
+            gr.update(value=_score_radio_value(3)),
+            gr.update(value=_camera_hint_before_start()),
+            st,
+            gr.update(value="重新攝影機辨識"),
+        )
+
+    top_key = max(emo_dict, key=lambda k: float(emo_dict[k]))
+    top_conf = float(emo_dict[top_key])
+
     if top_conf < CONF_TH:
-        show = st.stable_emo or "unknown"
+        annotated = _draw_guide_overlay(frame_rgb, True)
         return (
-            gr.update(value=frame_show, visible=True),
-            gr.update(value=_result_card_html(show, st.final_conf)),
-            gr.update(value=_score_radio_value(EMO_TO_SCORE.get(show, 3))),
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("已偵測到人臉，開始辨識情緒中…")),
+            gr.update(value=_score_radio_value(3)),
             gr.update(value=_camera_hint_before_start()),
             st,
             gr.update(value="重新攝影機辨識"),
         )
 
-    # === EMA smoothing ===
-    # DeepFace emotion dict 的 keys 順序可能固定，但我們這裡用單一 top_conf/top_key 也能做穩定
-    # 仍然保留你原本的「看整個分佈」的 smoothing：用 top_key/top_conf 建一個簡化 dict
-    # 為了兼容你原本寫法，我們用 7 類別做一個向量（缺的�� 0）
-    all_keys = ["angry", "disgust", "fear", "happy", "sad", "surprise", "neutral"]
-    vec = np.zeros((len(all_keys),), dtype=np.float32)
-    if top_key in all_keys:
-        vec[all_keys.index(top_key)] = float(top_conf)
+    keys = list(emo_dict.keys())
+    vec = np.array([float(emo_dict[k]) for k in keys], dtype=np.float32)
 
-    if st.ema_scores is None or st.emo_keys != all_keys:
+    if st.ema_scores is None or st.emo_keys != keys:
         st.ema_scores = vec
-        st.emo_keys = all_keys
+        st.emo_keys = keys
     else:
         st.ema_scores = EMA_ALPHA * vec + (1.0 - EMA_ALPHA) * st.ema_scores
 
     stable_idx = int(np.argmax(st.ema_scores))
-    stable_key = st.emo_keys[stable_idx] if st.emo_keys else "unknown"
+    stable_key = st.emo_keys[stable_idx]
     st.stable_emo = stable_key
 
-    # 延遲確認，避免跳動
     if st.candidate_emo != stable_key:
         st.candidate_emo = stable_key
         st.candidate_since = now
@@ -497,27 +528,25 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
         and (now - st.candidate_since) >= SWITCH_CONFIRM_SEC
     )
 
-    # ✅ 保留你原本「辨識完成後自動跳轉/停止」邏輯
     if confirmed and st.final_emo is None:
         st.final_emo = stable_key
-        st.final_conf = float(top_conf) if top_conf is not None else None
+        st.final_conf = top_conf
         st.final_score = EMO_TO_SCORE.get(stable_key, 3)
-
         st.running = False
         st.finished = True
-
         return (
-            gr.update(value=None, visible=False),
-            gr.update(value=_result_card_html(stable_key, st.final_conf)),
+            gr.update(visible=False),
+            gr.update(value=_result_card_html(stable_key, top_conf)),
             gr.update(value=_score_radio_value(EMO_TO_SCORE.get(stable_key, 3))),
             gr.update(value=_camera_hint_done()),
             st,
             gr.update(value="重新攝影機辨識"),
         )
 
+    annotated = _draw_guide_overlay(frame_rgb, True)
     return (
-        gr.update(value=frame_show, visible=True),
-        gr.update(value=_result_card_html(stable_key, float(top_conf))),
+        gr.update(visible=True, value=annotated),
+        gr.update(value=_result_card_html(stable_key, top_conf)),
         gr.update(value=_score_radio_value(EMO_TO_SCORE.get(stable_key, 3))),
         gr.update(value=_camera_hint_before_start()),
         st,
@@ -531,7 +560,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
 def on_restart(st: AppState):
     st = _reset_state_for_camera(st)
     return (
-        gr.update(value=None, visible=True),
+        gr.update(visible=True, value=None),  # ✅ 清掉上一張畫面
         gr.update(value=_hint_html("等待影像…（尚未開始串流）")),
         gr.update(value=_score_radio_value(3)),
         gr.update(value=_camera_hint_before_start()),
@@ -542,19 +571,17 @@ def on_restart(st: AppState):
 def on_stop(st: AppState):
     st.running = False
     st.finished = True
-
     if st.final_emo is None:
         return (
-            gr.update(value=None, visible=False),
+            gr.update(visible=False, value=None),
             gr.update(value=_hint_html("已停止辨識。")),
             gr.update(value=_score_radio_value(3)),
             gr.update(value=_camera_hint_stopped()),
             st,
             gr.update(value="重新攝影機辨識"),
         )
-
     return (
-        gr.update(value=None, visible=False),
+        gr.update(visible=False, value=None),
         gr.update(value=_result_card_html(st.final_emo, st.final_conf)),
         gr.update(value=_score_radio_value(EMO_TO_SCORE.get(st.final_emo, 3))),
         gr.update(value=_camera_hint_done()),
@@ -566,20 +593,19 @@ def on_stop(st: AppState):
 # =========================
 # Gradio UI
 # =========================
-css = ""
-
 if __name__ == "__main__":
     with gr.Blocks(title="Emotion Detector", css=css) as demo:
         st = gr.State(AppState())
 
         gr.Markdown("## 情緒辨識")
+
         hint = gr.HTML(_camera_hint_before_start())
 
         cam = gr.Image(
             sources=["webcam"],
             streaming=True,
             type="numpy",
-            label="攝影機畫面",
+            label="攝影機",
             visible=True,
         )