Update emotion.py

emotion.py

@@ -5,9 +5,8 @@ from typing import Optional, List, Tuple, Dict
 
 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
@@ -18,7 +17,7 @@ from deepface import DeepFace
 # =========================
 # 參數（可微調）
 # =========================
-ANALYZE_EVERY_SEC = 2.0          # ⭐ 建議 1.5~2.0：HF CPU 比較順
+ANALYZE_EVERY_SEC = 1.5          # ⭐ 建議 1.5~2.0：HF CPU 比較順
 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
@@ -29,7 +28,7 @@ UI_TARGET_W = 720                # ✅ 回傳畫面縮小寬度，提升順暢
 MP_MIN_DET_CONF = 0.3
 
 # 模型路徑（確認 repo 中模型真的在這裡）
-MP_MODEL_PATH = "blaze_face_short_range.tflite"
+MP_MODEL_PATH = "model/blaze_face_short_range.tflite"
 
 
 # =========================
@@ -248,7 +247,6 @@ def mp_tasks_detect_and_draw(frame_rgb: np.ndarray):
     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)
@@ -265,18 +263,14 @@ def mp_tasks_detect_and_draw(frame_rgb: np.ndarray):
     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)
-
-    # 白框
     cv2.rectangle(bgr, (lx, ly), (lx + bw, ly + bh), (255, 255, 255), 3)
 
-    # 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)
@@ -287,15 +281,19 @@ def mp_tasks_detect_and_draw(frame_rgb: np.ndarray):
 
 
 # =========================
-# 背景推論（避免卡頓）
+# 背景推論（避免卡頓） + ✅ 沒臉時清空
 # =========================
 _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 _clear_last_analyze():
+    """✅ 沒有人臉時，清空上一筆情緒，避免殘留顯示（你影片遇到的 bug）"""
+    global _LAST_ANALYZE
+    _LAST_ANALYZE = {"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:
@@ -311,7 +309,7 @@ def _analyze_worker():
                 img_path=face,
                 actions=["emotion"],
                 enforce_detection=False,
-                detector_backend="skip",  # ⭐ 關鍵：你已經裁好 ROI
+                detector_backend="skip",  # ⭐ 你已裁 ROI，別再偵測一次
             )
             if isinstance(r, list):
                 r = r[0]
@@ -322,52 +320,21 @@ def _analyze_worker():
                 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 _resize_letterbox_rgb(img: np.ndarray, size: int = 224) -> np.ndarray:
-    """
-    把 ROI 變成 size×size，保持比例，用黑邊補滿（letterbox）。
-    好處：不會把臉拉伸變形，對情緒分類通常更穩。
-    """
-    img = _to_rgb_uint8(img)
-    h, w = img.shape[:2]
-    if h <= 0 or w <= 0:
-        return img
-
-    scale = size / max(h, w)
-    nh, nw = int(h * scale), int(w * scale)
-    nh = max(1, nh)
-    nw = max(1, nw)
-
-    resized = cv2.resize(img, (nw, nh), interpolation=cv2.INTER_AREA)
-
-    out = np.zeros((size, size, 3), dtype=np.uint8)
-    y0 = (size - nh) // 2
-    x0 = (size - nw) // 2
-    out[y0:y0+nh, x0:x0+nw] = resized
-    return out
-
-
 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
 
-    # ✅ 這行是你要的：先把 ROI 縮到固定大小（224×224）
-    face_small = _resize_letterbox_rgb(face_roi, size=224)
-
     with _ANALYZE_LOCK:
-        _PENDING_FACE = face_small
+        _PENDING_FACE = face_roi
 
     if _ANALYZE_THREAD is None:
         _ANALYZE_THREAD = threading.Thread(target=_analyze_worker, daemon=True)
         _ANALYZE_THREAD.start()
 
 
-
 # =========================
 # state helper
 # =========================
@@ -386,6 +353,8 @@ def _reset_state_for_camera(st: AppState) -> AppState:
     st.final_emo = None
     st.final_conf = None
     st.final_score = 3
+
+    _clear_last_analyze()
     return st
 
 
@@ -426,18 +395,25 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
             gr.update(value="重新攝影機辨識"),
         )
 
-    # 統���格式
     frame_rgb = _to_rgb_uint8(frame_rgb)
 
-    # ✅ 偵測用原圖（穩）
+    # ✅ 偵測用原圖
     frame_drawn, bbox_px = mp_tasks_detect_and_draw(frame_rgb)
 
-    # ✅ 顯示用 zoom（你要保留 zoom 畫面）
+    # ✅ 顯示用 zoom
     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)
 
-    # 沒有人臉：不畫框/點 + 提示
+    # ✅ 沒有臉：清空殘留情緒 + 重置候選狀態
     if bbox_px is None:
+        _clear_last_analyze()
+
+        st.stable_emo = "unknown"
+        st.candidate_emo = None
+        st.candidate_since = None
+        st.ema_scores = None
+        st.emo_keys = []
+
         return (
             gr.update(value=frame_show, visible=True),
             gr.update(value=_hint_html("尚未偵測到人臉")),
@@ -449,7 +425,7 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
 
     now = time.time()
 
-    # 到分析時間：只排隊給背景 thread（不在 on_stream 裡跑重推論）
+    # 到分析時間：只排隊給背景 thread
     if now - st.last_analyze_t >= ANALYZE_EVERY_SEC:
         st.last_analyze_t = now
 
@@ -465,38 +441,35 @@ def on_stream(frame_rgb: np.ndarray, st: AppState):
 
         _enqueue_analyze(face_roi)
 
-    # 取「最後一次背景推論結果」
+    # ✅ Gate：只有「有臉」才讀取 _LAST_ANALYZE（避免殘留）
     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
+    # 尚未推論到結果：先顯示 unknown
     if top_conf is None:
         show = st.stable_emo or "unknown"
         return (
             gr.update(value=frame_show, visible=True),
-            gr.update(value=_result_card_html(show, st.final_conf)),
+            gr.update(value=_result_card_html(show, None)),
             gr.update(value=_score_radio_value(EMO_TO_SCORE.get(show, 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
+    # 低信心就不切換
     if top_conf < CONF_TH:
         show = st.stable_emo or "unknown"
         return (
             gr.update(value=frame_show, visible=True),
-            gr.update(value=_result_card_html(show, st.final_conf)),
+            gr.update(value=_result_card_html(show, None)),
             gr.update(value=_score_radio_value(EMO_TO_SCORE.get(show, 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（缺的填 0）
+    # === EMA smoothing（用 7 類別向量）===
     all_keys = ["angry", "disgust", "fear", "happy", "sad", "surprise", "neutral"]
     vec = np.zeros((len(all_keys),), dtype=np.float32)
     if top_key in all_keys:
@@ -525,7 +498,7 @@ 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
+    # ✅ 完成一次辨識後停止
     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