Update emotion.py

emotion.py

@@ -1,49 +1,150 @@
 import time
-import threading
 from dataclasses import dataclass, field
 from typing import Optional, Dict, Tuple, List
 
 import gradio as gr
 import numpy as np
 from PIL import Image, ImageDraw
+
 from deepface import DeepFace
 
+
+# ✅ 避免 css 未定義造成 NameError
+css = ""
+
+
+# =========================
+# 參數（HF Space 建議偏順）
+# =========================
+ANALYZE_EVERY_SEC = 1.4
+DOWNSAMPLE_W = 480
+EMA_ALPHA = 0.45
+CONF_TH = 40.0
+SWITCH_CONFIRM_SEC = 2.2
+IOU_TH = 0.25
+ZOOM_FACTOR = 1.25
+
+DETECTOR_BACKEND = "opencv"   # Space 上較順
+ALIGN_FACE = True
+
+MIN_FACE_AREA_RATIO = 0.04    # 人臉太小視為沒偵測到
+RECENT_FACE_SEC = 0.25        # ✅ 綠框快取時間：縮短避免空景綠框殘留
+
+
+# =========================
+# OpenCV 眼睛驗證（降低背景假臉）
+# =========================
+try:
+    import cv2
+    _HAS_CV2 = True
+    _EYE_CASCADE = cv2.CascadeClassifier(
+        cv2.data.haarcascades + "haarcascade_eye.xml"
+    )
+except Exception:
+    _HAS_CV2 = False
+    _EYE_CASCADE = None
+
+
+def _verify_face_has_eyes(face_rgb_uint8: np.ndarray) -> bool:
+    """
+    用眼睛 HaarCascade 做簡單驗證：
+    - 背景假臉通常偵測不到眼睛
+    - 真臉通常能偵測到 >= 1 個眼睛
+    """
+    if not _HAS_CV2 or _EYE_CASCADE is None:
+        # 沒有 cv2 就不做驗證（維持原行為）
+        return True
+
+    if face_rgb_uint8 is None or face_rgb_uint8.size == 0:
+        return False
+
+    # 影像太小就直接判 False
+    h, w = face_rgb_uint8.shape[:2]
+    if h < 40 or w < 40:
+        return False
+
+    gray = cv2.cvtColor(face_rgb_uint8, cv2.COLOR_RGB2GRAY)
+    gray = cv2.equalizeHist(gray)
+
+    # scaleFactor / minNeighbors 調得偏嚴格 → 更能擋掉背景假臉
+    eyes = _EYE_CASCADE.detectMultiScale(
+        gray,
+        scaleFactor=1.12,
+        minNeighbors=6,
+        minSize=(12, 12),
+    )
+    return len(eyes) >= 1
+
+
 # =========================
-# 1. 核心參數優化
+# 情緒中文 + 分數映射
 # =========================
-ANALYZE_EVERY_SEC = 1.0     # 辨識頻率：1秒1次 (降低 CPU 負荷)
-DOWNSAMPLE_W = 320          # 解析度降至 320px (大幅提升速度)
-EMA_ALPHA = 0.40            # 平滑係數
-CONF_TH = 35.0              # 信心值門檻
-SWITCH_CONFIRM_SEC = 2.0    # 情緒穩定多久才算最終結果
-ZOOM_FACTOR = 1.25          # 稍微放大臉部區域
-DETECTOR_BACKEND = "opencv" # 最速辨識器
-ALIGN_FACE = False          # 關閉對齊以節省資源
+EMO_ZH = {
+    "angry": "生氣",
+    "disgust": "厭惡",
+    "fear": "害怕",
+    "happy": "開心",
+    "sad": "難過",
+    "surprise": "驚訝",
+    "neutral": "平靜",
+    "unknown": "未知",
+}
+
+EMO_TO_SCORE = {
+    "angry": 1,
+    "disgust": 1,
+    "fear": 1,
+    "sad": 2,
+    "neutral": 3,
+    "surprise": 4,
+    "happy": 5,
+    "unknown": 3,
+}
+
+SCORE_LABELS = {
+    1: "1（心情差）",
+    2: "2（不太好）",
+    3: "3（普通）",
+    4: "4（不錯）",
+    5: "5（超棒）",
+}
+SCORE_CHOICES = [SCORE_LABELS[i] for i in [1, 2, 3, 4, 5]]
+
+def _score_radio_value(score: int) -> str:
+    score = int(score)
+    score = min(5, max(1, score))
+    return SCORE_LABELS[score]
 
-# 全域狀態：防止 DeepFace 重疊執行導致阻塞
-_is_processing = False
 
 # =========================
-# 2. 狀態管理
+# 狀態
 # =========================
 @dataclass
 class AppState:
     running: bool = True
     finished: bool = False
     last_analyze_t: float = 0.0
+
     track_bbox: Optional[Tuple[int, int, int, int]] = None
+
     ema_scores: Optional[np.ndarray] = None
     emo_keys: List[str] = field(default_factory=list)
+
     stable_emo: str = "unknown"
     candidate_emo: Optional[str] = None
     candidate_since: Optional[float] = None
+
     final_emo: Optional[str] = None
     final_conf: Optional[float] = None
+    final_score: int = 3
+
+    face_ok: bool = False
     last_face_t: float = 0.0
     face_in_box: bool = False
 
+
 # =========================
-# 3. UI 輔助函數 (修正 AttributeError)
+# UI HTML
 # =========================
 def _hint_html(msg: str) -> str:
     return f"""
@@ -52,178 +153,536 @@ def _hint_html(msg: str) -> str:
     </div>
     """
 
+def _camera_hint_before_start() -> str:
+    return _hint_html(
+        "請先按「Click to access webcam / 允許」開啟攝影機；"
+        "接著，請按「錄製（Record）」開始。"
+    )
+
+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 _result_card_html(emo_key: str, conf: Optional[float]) -> str:
-    EMO_ZH = {"angry": "生氣", "disgust": "厭惡", "fear": "害怕", "happy": "開心", "sad": "難過", "surprise": "驚訝", "neutral": "平靜", "unknown": "未知"}
-    zh = EMO_ZH.get(emo_key, "分析中...")
-    score_map = {"angry": 1, "disgust": 1, "fear": 1, "sad": 2, "neutral": 3, "surprise": 4, "happy": 5, "unknown": 3}
-    score = score_map.get(emo_key, 3)
-    conf_txt = f"{conf:.1f}%" if conf else "--"
+    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:#666;margin-bottom:8px;">即時辨識結果</div>
-      <div style="font-size:42px;font-weight:800;line-height:1.1;margin-bottom:14px;">{zh}</div>
-      <div style="font-size:16px;margin-bottom:6px;">心情分數：{score}</div>
-      <div style="font-size:14px;color:#999;">信心值：{conf_txt}</div>
+      <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_to_label(score_int: int) -> str:
-    labels = {1: "1（心情差）", 2: "2（不太好）", 3: "3（普通）", 4: "4（不錯）", 5: "5（超棒）"}
-    return labels.get(score_int, "3（普通）")
 
 # =========================
-# 4. 影像處理工具
+# 影像工具
 # =========================
 def _downsample_rgb(frame: np.ndarray, target_w: int) -> np.ndarray:
-    if frame is None: return None
+    if frame is None:
+        return None
     h, w = frame.shape[:2]
-    if w <= target_w: return frame
+    if w <= target_w:
+        return frame
     scale = target_w / w
-    img = Image.fromarray(frame).resize((target_w, int(h * scale)), Image.NEAREST)
+    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: return frame
+    if frame is None or zoom <= 1.0:
+        return frame
     h, w = frame.shape[:2]
-    nw, nh = int(w / zoom), int(h / zoom)
-    sx, sy = (w - nw) // 2, (h - nh) // 2
-    return frame[sy:sy + nh, sx:sx + nw]
+    new_w = int(w / zoom)
+    new_h = int(h / zoom)
+    start_x = (w - new_w) // 2
+    start_y = (h - new_h) // 2
+    return frame[start_y:start_y + new_h, start_x:start_x + new_w]
+
+
+# =========================
+# 引導框（固定位置）
+# =========================
+GUIDE_W_RATIO = 0.60
+GUIDE_H_RATIO = 0.72
+GUIDE_Y_OFFSET = 0.00
+
+def _guide_box_xyxy(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, green: bool, msg: str) -> np.ndarray:
+    img = Image.fromarray(frame_rgb)
+    draw = ImageDraw.Draw(img)
+
+    x1, y1, x2, y2 = _guide_box_xyxy(frame_rgb)
+    color = (0, 200, 0) if green else (220, 0, 0)
+
+    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)
+
+    tx = x1
+    ty = max(0, y1 - int(thickness * 7))
+    draw.rectangle([tx, ty, min(w - 1, tx + 520), ty + 34], fill=(255, 255, 255))
+    draw.text((tx + 8, ty + 7), msg, fill=(0, 0, 0))
+
+    return np.array(img)
+
 
 # =========================
-# 5. 核心串流邏輯 (防止阻塞)
+# Tracking / IoU
+# =========================
+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]
+
+
+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.face_ok = False
+    st.last_face_t = 0.0
+    st.face_in_box = False
+    return st
+
+
+def _is_face_big_enough(bbox_xywh: Tuple[float, float, float, float], frame_hw: Tuple[int, int]) -> bool:
+    fh, fw = frame_hw
+    x, y, w, h = bbox_xywh
+    area_ratio = (w * h) / max(1.0, float(fw * fh))
+    return area_ratio >= MIN_FACE_AREA_RATIO
+
+
+# =========================
+# Camera streaming（含紅/綠框 + 眼睛驗證）
 # =========================
 def on_stream(frame_rgb: np.ndarray, st: AppState):
-    global _is_processing
+    if st.finished or (not st.running):
+        cam_visible = False
+        if st.final_emo:
+            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(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"),
+        )
+
+    if frame_rgb is None:
+        return (
+            gr.update(visible=True, value=None),
+            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()),
+            st,
+            gr.update(value="重新攝影機辨識"),
+        )
+
+    frame_rgb = _zoom_center(frame_rgb, ZOOM_FACTOR)
     now = time.time()
 
-    # 如果已經完成或沒影像，直接結束
-    if st.finished or not st.running or frame_rgb is None:
-        return (gr.update(), gr.update(), gr.update(), gr.update(), st)
+    # ✅ 快取只留很短時間（避免空景也綠）
+    recent_face = st.face_ok and st.face_in_box and (now - st.last_face_t) < RECENT_FACE_SEC
+    msg_fast = "已偵測到人臉，開始辨識情緒中…" if recent_face else "未偵測到人臉，請把臉放進框內"
+    annotated_fast = _draw_guide_overlay(frame_rgb, green=recent_face, msg=msg_fast)
 
-    # A. 基礎影像處理 (輕量級，每影格都做)
-    processed_frame = _zoom_center(frame_rgb, ZOOM_FACTOR)
+    if now - st.last_analyze_t < ANALYZE_EVERY_SEC:
+        if recent_face:
+            show = st.stable_emo or "unknown"
+            result_html = _result_card_html(show, st.final_conf)
+            mood_val = _score_radio_value(EMO_TO_SCORE.get(show, 3))
+        else:
+            result_html = _hint_html("未偵測到人臉，請把臉放進框內（紅框）")
+            mood_val = _score_radio_value(3)
 
-    # B. 判斷是否要進行 DeepFace 分析 (防止堆積)
-    if _is_processing or (now - st.last_analyze_t < ANALYZE_EVERY_SEC):
-        # 忙碌中或時間未到，只回傳影像，確保 UI 不會卡死
-        return (processed_frame, gr.update(), gr.update(), gr.update(), st)
+        return (
+            gr.update(visible=True, value=annotated_fast),
+            gr.update(value=result_html),
+            gr.update(value=mood_val),
+            gr.update(value=_camera_hint_before_start()),
+            st,
+            gr.update(value="重新攝影機辨識"),
+        )
+
+    st.last_analyze_t = now
+    small = _downsample_rgb(frame_rgb, DOWNSAMPLE_W)
 
-    # C. 開始分析 (加鎖)
-    _is_processing = True
     try:
-        st.last_analyze_t = now
-        small_img = _downsample_rgb(processed_frame, DOWNSAMPLE_W)
-        
-        # 使用 DeepFace 分析 (enforce_detection=False 避免沒臉時噴 Error)
-        results = DeepFace.analyze(
-            img_path=small_img,
-            actions=['emotion'],
+        faces = DeepFace.extract_faces(
+            img_path=small,
             detector_backend=DETECTOR_BACKEND,
             enforce_detection=False,
-            align=ALIGN_FACE
-        )
-        
-        if results and len(results) > 0:
-            res = results[0]
-            # 只有在真的偵測到臉的情況下才更新 (OpenCV 信心值較低)
-            if res.get('face_confidence', 0) > 0.3:
-                st.last_face_t = now
-                emo_dict = res['emotion']
-                top_key = max(emo_dict, key=emo_dict.get)
-                
-                # EMA 平滑
-                keys = list(emo_dict.keys())
-                vec = np.array([float(emo_dict[k]) for k in keys])
-                if st.ema_scores is None:
-                    st.ema_scores, st.emo_keys = vec, keys
-                else:
-                    st.ema_scores = EMA_ALPHA * vec + (1.0 - EMA_ALPHA) * st.ema_scores
-                
-                stable_key = st.emo_keys[np.argmax(st.ema_scores)]
-                st.stable_emo = stable_key
-
-                # 確認邏輯
-                if st.candidate_emo != stable_key:
-                    st.candidate_emo, st.candidate_since = stable_key, now
-                elif st.candidate_since and (now - st.candidate_since) >= SWITCH_CONFIRM_SEC:
-                    st.final_emo, st.final_conf = stable_key, float(emo_dict[stable_key])
-                    st.finished = True
-                
-                score_val = {"angry": 1, "disgust": 1, "fear": 1, "sad": 2, "neutral": 3, "surprise": 4, "happy": 5, "unknown": 3}.get(stable_key, 3)
-                
-                return (
-                    processed_frame, 
-                    _result_card_html(stable_key, float(emo_dict[top_key])),
-                    _score_to_label(score_val),
-                    _hint_html("正在辨識情緒，請保持表情穩定..."),
-                    st
-                )
-
-    except Exception as e:
-        print(f"DeepFace 分析異常: {e}")
-    finally:
-        _is_processing = False
-
-    return (processed_frame, gr.update(), gr.update(), gr.update(), st)
-
-# =========================
-# 6. Gradio 介面佈局
-# =========================
-with gr.Blocks(title="Emotion AI Detector", css=".gradio-container {background-color: #f9f9f9}") as demo:
-    st_state = gr.State(AppState())
-    
-    gr.Markdown("# 🎭 AI 情緒即時偵測系統")
-    
-    with gr.Row():
-        with gr.Column(scale=2):
-            # 串流組件
-            input_cam = gr.Image(sources=["webcam"], streaming=True, type="numpy", label="即時視訊")
-            
-            with gr.Row():
-                btn_reset = gr.Button("🔄 重新辨識", variant="secondary")
-                btn_stop = gr.Button("⏹ 停止", variant="stop")
-        
-        with gr.Column(scale=1):
-            # 結果顯示組件
-            hint_box = gr.HTML(_hint_html("請開啟攝影機，並按「Record」開始辨識。"))
-            result_box = gr.HTML(_result_card_html("unknown", None))
-            mood_radio = gr.Radio(
-                choices=["1（心情差）", "2（不太好）", "3（普通）", "4（不錯）", "5（超棒）"],
-                label="系統判定心情等級",
-                interactive=False,
-                value="3（普通）"
-            )
+            align=ALIGN_FACE,
+        )
+    except Exception:
+        faces = []
+
+    pick, bbox = _pick_face_by_tracking(faces, st.track_bbox)
+
+    # 沒臉
+    if pick is None or bbox is None or pick.get("face") is None:
+        st.track_bbox = None
+        st.face_ok = False
+        st.face_in_box = False
+        st.last_face_t = now
+        st.stable_emo = "unknown"   # ✅ 清掉殘留情緒
+
+        annotated = _draw_guide_overlay(frame_rgb, green=False, msg="未偵測到人臉，請把臉放進框內")
+        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（紅框）")),
+            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"),
+        )
+
+    # bbox small -> full
+    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)
+
+    # 太小
+    if not _is_face_big_enough(bbox_full, (fh, fw)):
+        st.track_bbox = None
+        st.face_ok = False
+        st.face_in_box = False
+        st.last_face_t = now
+        st.stable_emo = "unknown"
+
+        annotated = _draw_guide_overlay(frame_rgb, green=False, msg="人臉太小/太遠，請靠近鏡頭並放進框內")
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("人臉太小/太遠，請靠近鏡頭並放進框內（紅框）")),
+            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_xyxy(frame_rgb)
+    in_box = _bbox_center_in_box(bbox_full, guide)
+
+    st.track_bbox = bbox
+    st.last_face_t = now
+    st.face_ok = True
+    st.face_in_box = in_box
+
+    if not in_box:
+        st.stable_emo = "unknown"
+        annotated = _draw_guide_overlay(frame_rgb, green=False, msg="已偵測到人臉，請把臉移到框內")
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("已偵測到人臉，但請把臉移到框內（紅框）")),
+            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"),
+        )
+
+    # ✅ 眼睛驗證（關鍵）
+    face = pick["face"]
+    if face.dtype != np.uint8:
+        face = np.clip(face * 255.0, 0, 255).astype(np.uint8)
+
+    if not _verify_face_has_eyes(face):
+        # 當作沒臉：避免背景假臉
+        st.track_bbox = None
+        st.face_ok = False
+        st.face_in_box = False
+        st.last_face_t = now
+        st.stable_emo = "unknown"
+
+        annotated = _draw_guide_overlay(frame_rgb, green=False, msg="未偵測到真實人臉（找不到眼睛），請對準鏡頭")
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("未偵測到真實人臉（找不到眼睛），請把臉放進框內（紅框）")),
+            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"),
+        )
+
+    # ✅ 到這裡才算「真的有人臉」→ 綠框 + 情緒分析
+    annotated = _draw_guide_overlay(frame_rgb, green=True, msg="已偵測到人臉，開始辨識情緒中…")
+
+    try:
+        result = DeepFace.analyze(
+            img_path=face,
+            actions=["emotion"],
+            enforce_detection=False,
+        )
+    except Exception:
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("已偵測到人臉，辨識中…（分析暫時失敗，請保持正面/光線亮一點）")),
+            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:
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("已偵測到人臉，辨識中…")),
+            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])
 
-    # --- 事件綁定 ---
-    
-    # 串流主邏輯: queue=False 與 show_progress="hidden" 是流暢關鍵
-    input_cam.stream(
-        fn=on_stream,
-        inputs=[input_cam, st_state],
-        outputs=[input_cam, result_box, mood_radio, hint_box, st_state],
-        show_progress="hidden",
-        queue=False
+    if top_conf < CONF_TH:
+        return (
+            gr.update(visible=True, value=annotated),
+            gr.update(value=_hint_html("已偵測到人臉，辨識中…（信心偏低，請面向鏡頭/光線亮一點）")),
+            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"),
+        )
+
+    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 != keys:
+        st.ema_scores = vec
+        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]
+    st.stable_emo = stable_key
+
+    if st.candidate_emo != stable_key:
+        st.candidate_emo = stable_key
+        st.candidate_since = now
+    else:
+        if st.candidate_since is None:
+            st.candidate_since = now
+
+    confirmed = (
+        st.candidate_since is not None
+        and (now - st.candidate_since) >= SWITCH_CONFIRM_SEC
     )
 
-    # 重新辨識按鈕
-    def restart_app():
-        return (None, _result_card_html("unknown", None), "3（普通）", _hint_html("系統已重置，請重新開始。"), AppState())
-    
-    btn_reset.click(
-        fn=restart_app,
-        outputs=[input_cam, result_box, mood_radio, hint_box, st_state]
+    if confirmed and st.final_emo is None:
+        st.final_emo = stable_key
+        st.final_conf = top_conf
+        st.final_score = EMO_TO_SCORE.get(stable_key, 3)
+        st.running = False
+        st.finished = True
+
+        return (
+            gr.update(visible=False, value=None),
+            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"),
+        )
+
+    return (
+        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,
+        gr.update(value="重新攝影機辨識"),
     )
 
-    # 停止按鈕
-    def stop_app(curr_st: AppState):
-        curr_st.running = False
-        return _hint_html("偵測已停止。"), curr_st
-    
-    btn_stop.click(fn=stop_app, inputs=[st_state], outputs=[hint_box, st_state])
 
 # =========================
-# 7. 啟動伺服器
+# Buttons
+# =========================
+def on_restart(st: AppState):
+    st = _reset_state_for_camera(st)
+    return (
+        gr.update(visible=True, value=None),
+        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()),
+        st,
+        gr.update(value="重新攝影機辨識"),
+    )
+
+def on_stop(st: AppState):
+    st.running = False
+    st.finished = True
+    if st.final_emo is None:
+        return (
+            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(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()),
+        st,
+        gr.update(value="重新攝影機辨識"),
+    )
+
+
+# =========================
+# Gradio UI
 # =========================
 if __name__ == "__main__":
-    # 使用 share=False 減少網路延遲造成的卡頓
-    demo.launch(server_name="0.0.0.0", server_port=7860, share=False)
+    with gr.Blocks(title="Emotion Detector", css=css) as demo:
+        st = gr.State(AppState())
+
+        gr.Markdown("## 第一步：測測你的心情能量")
+        gr.Markdown("讓 AI 看看你的表情，自動幫你決定心情分數！（辨識完畢會自動跳轉）")
+
+        hint = gr.HTML(_camera_hint_before_start())
+
+        cam = gr.Image(
+            sources=["webcam"],
+            streaming=True,
+            type="numpy",
+            label="攝影機畫面",
+            visible=True,
+        )
+
+        with gr.Row():
+            btn_restart = gr.Button("重新攝影機辨識", variant="primary")
+            btn_stop = gr.Button("停止", variant="secondary")
+
+        gr.Markdown("### 辨識結果")
+        result = gr.HTML(_hint_html("等待影像…（尚未開始串流）"))
+
+        gr.Markdown("### 心情分數（系統判定）")
+        mood = gr.Radio(
+            choices=SCORE_CHOICES,
+            value=_score_radio_value(3),
+            label="心情分數",
+            interactive=False,
+        )
+
+        cam.stream(
+            fn=on_stream,
+            inputs=[cam, st],
+            outputs=[cam, result, mood, hint, st, btn_restart],
+            show_progress="minimal",
+        )
+
+        btn_restart.click(
+            fn=on_restart,
+            inputs=[st],
+            outputs=[cam, result, mood, hint, st, btn_restart],
+            show_progress="minimal",
+        )
+
+        btn_stop.click(
+            fn=on_stop,
+            inputs=[st],
+            outputs=[cam, result, mood, hint, st, btn_restart],
+            show_progress="minimal",
+        )
+
+    demo.launch(server_name="0.0.0.0", server_port=7860, share=False)