face_analyzer.py

import os
import random
import time
from typing import List, Dict, Any, Optional

import cv2
import numpy as np

import config
from config import logger, MODELS_PATH, OUTPUT_DIR, DEEPFACE_AVAILABLE, \
    YOLO_AVAILABLE
from facial_analyzer import FacialFeatureAnalyzer
from models import ModelType
from utils import save_image_high_quality, get_date_subfolder

if DEEPFACE_AVAILABLE:
    from deepface import DeepFace

# 可选导入 YOLO
if YOLO_AVAILABLE:
    try:
        from ultralytics import YOLO

        YOLO_AVAILABLE = True
    except ImportError:
        YOLO_AVAILABLE = False
        YOLO = None
        print("Warning: ENABLE_YOLO=true but ultralytics not available")


class EnhancedFaceAnalyzer:
    """增强版人脸分析器 - 支持混合模型"""

    def __init__(self, models_dir: str = MODELS_PATH):
        """
        初始化人脸分析器
        :param models_dir: 模型文件目录
        """
        start_time = time.perf_counter()
        self.models_dir = models_dir
        self.MODEL_MEAN_VALUES = (104, 117, 123)
        self.age_list = [
            "(0-2)",
            "(4-6)",
            "(8-12)",
            "(15-20)",
            "(25-32)",
            "(38-43)",
            "(48-53)",
            "(60-100)",
        ]
        self.gender_list = ["Male", "Female"]
        # 性别对应的颜色 (BGR格式)
        self.gender_colors = {
            "Male": (255, 165, 0),  # 橙色 Orange
            "Female": (255, 0, 255),  # 洋红 Magenta / Fuchsia
        }

        # 初始化五官分析器
        self.facial_analyzer = FacialFeatureAnalyzer()
        # 加载HowCuteAmI模型
        self._load_howcuteami_models()
        # 加载YOLOv人脸检测模型
        self._load_yolo_model()

        # 预热模型（可选，通过配置开关）
        if getattr(config, "ENABLE_WARMUP", False):
            self._warmup_models()

        init_time = time.perf_counter() - start_time
        logger.info(f"EnhancedFaceAnalyzer initialized successfully, time: {init_time:.3f}s")

    def _cap_conf(self, value: float) -> float:
        """将置信度限制在 [0, 0.9999] 并保留4位小数。"""
        try:
            v = float(value if value is not None else 0.0)
        except Exception:
            v = 0.0
        if v >= 1.0:
            v = 0.9999
        if v < 0.0:
            v = 0.0
        return round(v, 4)

    def _adjust_beauty_score(self, score: float) -> float:
        try:
            if not config.BEAUTY_ADJUST_ENABLED:
                return score
            # 读取提分区间与力度
            low = float(getattr(config, "BEAUTY_ADJUST_MIN", 6.0))
            high = float(getattr(config, "BEAUTY_ADJUST_MAX", getattr(config, "BEAUTY_ADJUST_THRESHOLD", 8.0)))
            gamma = float(getattr(config, "BEAUTY_ADJUST_GAMMA", 0.3))
            gamma = max(0.0001, min(1.0, gamma))

            # 区间有效性保护
            if not (0.0 <= low < high <= 10.0):
                return score

            # 低于下限不提分，区间内提向上限，高于上限不变
            if score < low:
                return score
            if score < high:
                # 向上限 high 进行温和靠拢：adjusted = high - gamma * (high - score)
                adjusted = high - gamma * (high - score)
                adjusted = round(min(10.0, max(0.0, adjusted)), 1)
                try:
                    logger.info(
                        f"beauty_score adjusted: original={score:.1f} -> adjusted={adjusted:.1f} "
                        f"(range=[{low:.1f},{high:.1f}], gamma={gamma:.3f})"
                    )
                except Exception:
                    pass
                return adjusted
            return score
        except Exception:
            return score

    def _load_yolo_model(self):
        """加载YOLOv人脸检测模型"""
        self.yolo_model = None
        if config.YOLO_AVAILABLE:
            try:
                # 尝试加载本地YOLOv人脸模型
                yolo_face_path = os.path.join(self.models_dir, config.YOLO_MODEL)

                if os.path.exists(yolo_face_path):
                    self.yolo_model = YOLO(yolo_face_path)
                    logger.info(f"Local YOLO face model loaded successfully: {yolo_face_path}")
                else:
                    # 如果本地没有，尝试在线下载（第一次使用时）
                    logger.info("Local YOLO face model does not exist, attempting to download...")
                    try:
                        # 检查是否是yolov8，使用相应的模型
                        model_name = "yolov11n-face.pt"  # 默认使用yolov8n
                        self.yolo_model = YOLO(model_name)
                        logger.info(
                            f"YOLOv8 general model loaded successfully (detecting 'person' class as face regions)"
                        )
                    except Exception as e:
                        logger.warning(f"YOLOv model download failed: {e}")

            except Exception as e:
                logger.error(f"YOLOv model loading failed: {e}")
        else:
            logger.warning("ultralytics not installed, cannot use YOLOv")

    def _load_howcuteami_models(self):
        """加载HowCuteAmI深度学习模型"""
        try:
            # 人脸检测模型
            face_proto = os.path.join(self.models_dir, "opencv_face_detector.pbtxt")
            face_model = os.path.join(self.models_dir, "opencv_face_detector_uint8.pb")
            self.face_net = cv2.dnn.readNet(face_model, face_proto)

            # 年龄预测模型
            age_proto = os.path.join(self.models_dir, "age_googlenet.prototxt")
            age_model = os.path.join(self.models_dir, "age_googlenet.caffemodel")
            self.age_net = cv2.dnn.readNet(age_model, age_proto)

            # 性别预测模型
            gender_proto = os.path.join(self.models_dir, "gender_googlenet.prototxt")
            gender_model = os.path.join(self.models_dir, "gender_googlenet.caffemodel")
            self.gender_net = cv2.dnn.readNet(gender_model, gender_proto)

            # 颜值预测模型
            beauty_proto = os.path.join(self.models_dir, "beauty_resnet.prototxt")
            beauty_model = os.path.join(self.models_dir, "beauty_resnet.caffemodel")
            self.beauty_net = cv2.dnn.readNet(beauty_model, beauty_proto)

            logger.info("HowCuteAmI model loaded successfully!")

        except Exception as e:
            logger.error(f"HowCuteAmI model loading failed: {e}")
            raise e

    # 人脸检测方法
    def _detect_faces(
        self, frame: np.ndarray, conf_threshold: float = config.FACE_CONFIDENCE
    ) -> List[List[int]]:
        """
        使用YOLO进行人脸检测，如果失败则回退到OpenCV DNN
        """
        # 优先使用YOLO
        face_boxes = []
        if self.yolo_model is not None:
            try:
                results = self.yolo_model(frame, conf=conf_threshold, verbose=False)
                for result in results:
                    boxes = result.boxes
                    if boxes is not None:
                        for box in boxes:
                            # 检查类别ID (如果是专门的人脸模型，通常是0；如果是通用模型，person类别通常是0)
                            class_id = int(box.cls[0])
                            # 获取边界框坐标 (xyxy格式)
                            x1, y1, x2, y2 = box.xyxy[0].cpu().numpy().astype(int)
                            confidence = float(box.conf[0])
                            logger.info(
                                f"detect class_id={class_id}, confidence={confidence}"
                            )
                            # 基本边界检查
                            frame_height, frame_width = frame.shape[:2]
                            x1 = max(0, int(x1))
                            y1 = max(0, int(y1))
                            x2 = min(frame_width, int(x2))
                            y2 = min(frame_height, int(y2))

                            # 过滤太小的检测框
                            width, height = x2 - x1, y2 - y1
                            if (
                                width > 30 and height > 30
                            ):  # YOLO通常检测精度更高，可以稍微提高最小尺寸
                                # 如果使用通用模型检测person，需要进一步过滤头部区域
                                if self._is_likely_face_region(x1, y1, x2, y2, frame):
                                    face_boxes.append(self._scale_box([x1, y1, x2, y2]))
                logger.info(
                    f"YOLO detected {len(face_boxes)} faces, conf_threshold={conf_threshold}"
                )
                if face_boxes:  # 如果YOLO检测到了人脸，直接返回
                    return face_boxes

            except Exception as e:
                logger.warning(f"YOLO detection failed, falling back to OpenCV DNN: {e}")
                return self._detect_faces_opencv_fallback(frame, conf_threshold)

        return face_boxes

    def _is_likely_face_region(
        self, x1: int, y1: int, x2: int, y2: int, frame: np.ndarray
    ) -> bool:
        """
        判断检测区域是否可能是人脸区域（当使用通用YOLO模型时）
        """
        width, height = x2 - x1, y2 - y1

        # 长宽比检查 - 人脸/头部通常接近正方形
        aspect_ratio = width / height
        if not (0.6 <= aspect_ratio <= 1.6):
            return False

        # 位置检查 - 人脸通常在图像上半部分（简单启发式）
        frame_height = frame.shape[0]
        center_y = (y1 + y2) / 2
        if center_y > frame_height * 0.8:  # 如果中心点在图像下方80%以下，可能不是人脸
            return False

        # 尺寸检查 - 不应该占据整个图像
        frame_width, frame_height = frame.shape[1], frame.shape[0]
        if width > frame_width * 0.8 or height > frame_height * 0.8:
            return False

        return True

    def _detect_faces_opencv_fallback(
        self, frame: np.ndarray, conf_threshold: float = 0.5
    ) -> List[List[int]]:
        """
        优化版人脸检测 - 支持多尺度检测和小人脸识别
        """
        frame_height, frame_width = frame.shape[:2]
        all_boxes = []

        # 多尺度检测配置 - 从小到大，更好地检测不同大小的人脸
        detection_configs = [
            {"size": (300, 300), "threshold": conf_threshold},
            {
                "size": (416, 416),
                "threshold": max(0.3, conf_threshold - 0.2),
            },  # 对大尺度降低阈值
            {
                "size": (512, 512),
                "threshold": max(0.25, conf_threshold - 0.25),
            },  # 进一步降低阈值检测小脸
        ]
        logger.info(f"Detecting faces using opencv, conf_threshold={conf_threshold}")
        for config in detection_configs:
            try:
                # 图像预处理 - 增强对比度有助于小人脸检测
                processed_frame = cv2.convertScaleAbs(frame, alpha=1.1, beta=10)

                blob = cv2.dnn.blobFromImage(
                    processed_frame, 1.0, config["size"], [104, 117, 123], True, False
                )
                self.face_net.setInput(blob)
                detections = self.face_net.forward()

                # 提取检测结果
                for i in range(detections.shape[2]):
                    confidence = detections[0, 0, i, 2]
                    if confidence > config["threshold"]:
                        x1 = int(detections[0, 0, i, 3] * frame_width)
                        y1 = int(detections[0, 0, i, 4] * frame_height)
                        x2 = int(detections[0, 0, i, 5] * frame_width)
                        y2 = int(detections[0, 0, i, 6] * frame_height)

                        # 基本边界检查
                        x1, y1 = max(0, x1), max(0, y1)
                        x2, y2 = min(frame_width, x2), min(frame_height, y2)

                        # 过滤太小或不合理的检测框
                        width, height = x2 - x1, y2 - y1
                        if (
                            width > 20
                            and height > 20
                            and width < frame_width * 0.8
                            and height < frame_height * 0.8
                        ):
                            # 长宽比检查 - 人脸通常接近正方形
                            aspect_ratio = width / height
                            if 0.6 <= aspect_ratio <= 1.8:  # 允许一定的椭圆形变
                                all_boxes.append(
                                    {
                                        "box": [x1, y1, x2, y2],
                                        "confidence": confidence,
                                        "area": width * height,
                                    }
                                )
            except Exception as e:
                logger.warning(f"Scale {config['size']} detection failed: {e}")
                continue

        # 如果没有检测到任何人脸，尝试更宽松的条件
        if not all_boxes:
            logger.info("No faces detected, trying more relaxed detection conditions...")
            try:
                # 最后一次尝试：最低阈值 + 图像增强
                enhanced_frame = cv2.equalizeHist(
                    cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY)
                )
                enhanced_frame = cv2.cvtColor(enhanced_frame, cv2.COLOR_GRAY2BGR)

                blob = cv2.dnn.blobFromImage(
                    enhanced_frame, 1.0, (300, 300), [104, 117, 123], True, False
                )
                self.face_net.setInput(blob)
                detections = self.face_net.forward()

                for i in range(detections.shape[2]):
                    confidence = detections[0, 0, i, 2]
                    if confidence > 0.15:  # 非常低的阈值
                        x1 = int(detections[0, 0, i, 3] * frame_width)
                        y1 = int(detections[0, 0, i, 4] * frame_height)
                        x2 = int(detections[0, 0, i, 5] * frame_width)
                        y2 = int(detections[0, 0, i, 6] * frame_height)

                        x1, y1 = max(0, x1), max(0, y1)
                        x2, y2 = min(frame_width, x2), min(frame_height, y2)

                        width, height = x2 - x1, y2 - y1
                        if width > 15 and height > 15:  # 更小的最小尺寸
                            aspect_ratio = width / height
                            if 0.5 <= aspect_ratio <= 2.0:  # 更宽松的长宽比
                                all_boxes.append(
                                    {
                                        "box": [x1, y1, x2, y2],
                                        "confidence": confidence,
                                        "area": width * height,
                                    }
                                )
            except Exception as e:
                logger.warning(f"Relaxed condition detection also failed: {e}")

        # NMS (非极大值抑制) 去除重复检测
        if all_boxes:
            final_boxes = self._apply_nms(all_boxes, overlap_threshold=0.4)
            return [self._scale_box(box["box"]) for box in final_boxes]

        return []

    def _apply_nms(
        self, detections: List[Dict], overlap_threshold: float = 0.4
    ) -> List[Dict]:
        """
        非极大值抑制，去除重复的检测框
        """
        if not detections:
            return []

        # 按置信度排序
        detections.sort(key=lambda x: x["confidence"], reverse=True)

        keep = []
        while detections:
            # 保留置信度最高的
            best = detections.pop(0)
            keep.append(best)

            # 移除与最佳检测重叠度高的其他检测
            remaining = []
            for det in detections:
                if self._calculate_iou(best["box"], det["box"]) < overlap_threshold:
                    remaining.append(det)
            detections = remaining

        return keep

    def _calculate_iou(self, box1: List[int], box2: List[int]) -> float:
        """
        计算两个边界框的IoU (交并比)
        """
        x1_1, y1_1, x2_1, y2_1 = box1
        x1_2, y1_2, x2_2, y2_2 = box2

        # 计算交集
        x1_i = max(x1_1, x1_2)
        y1_i = max(y1_1, y1_2)
        x2_i = min(x2_1, x2_2)
        y2_i = min(y2_1, y2_2)

        if x2_i <= x1_i or y2_i <= y1_i:
            return 0.0

        intersection = (x2_i - x1_i) * (y2_i - y1_i)

        # 计算并集
        area1 = (x2_1 - x1_1) * (y2_1 - y1_1)
        area2 = (x2_2 - x1_2) * (y2_2 - y1_2)
        union = area1 + area2 - intersection

        return intersection / union if union > 0 else 0.0

    def _scale_box(self, box: List[int]) -> List[int]:
        """将矩形框缩放为正方形"""
        width = box[2] - box[0]
        height = box[3] - box[1]
        maximum = max(width, height)
        dx = int((maximum - width) / 2)
        dy = int((maximum - height) / 2)

        return [box[0] - dx, box[1] - dy, box[2] + dx, box[3] + dy]

    def _crop_face(self, image: np.ndarray, box: List[int]) -> np.ndarray:
        """裁剪人脸区域"""
        x1, y1, x2, y2 = box
        h, w = image.shape[:2]
        x1 = max(0, x1)
        y1 = max(0, y1)
        x2 = min(w, x2)
        y2 = min(h, y2)
        return image[y1:y2, x1:x2]

    def _predict_beauty_gender_with_howcuteami(
        self, face: np.ndarray
    ) -> Dict[str, Any]:
        """使用HowCuteAmI模型预测颜值和性别"""
        try:
            blob = cv2.dnn.blobFromImage(
                face, 1.0, (224, 224), self.MODEL_MEAN_VALUES, swapRB=False
            )

            # 性别预测
            self.gender_net.setInput(blob)
            gender_preds = self.gender_net.forward()
            gender = self.gender_list[gender_preds[0].argmax()]
            gender_confidence = float(np.max(gender_preds[0]))
            gender_confidence = self._cap_conf(gender_confidence)
            # 年龄预测
            self.age_net.setInput(blob)
            age_preds = self.age_net.forward()
            age = self.age_list[age_preds[0].argmax()]
            age_confidence = float(np.max(age_preds[0]))
            # 颜值预测
            blob_beauty = cv2.dnn.blobFromImage(
                face, 1.0 / 255, (224, 224), self.MODEL_MEAN_VALUES, swapRB=False
            )
            self.beauty_net.setInput(blob_beauty)
            beauty_preds = self.beauty_net.forward()
            beauty_score = round(float(2.0 * np.sum(beauty_preds[0])), 1)
            beauty_score = min(10.0, max(0.0, beauty_score))
            beauty_score = self._adjust_beauty_score(beauty_score)
            raw_score = float(np.sum(beauty_preds[0]))

            return {
                "age": age,
                "age_confidence": round(age_confidence, 4),
                "gender": gender,
                "gender_confidence": gender_confidence,
                "beauty_score": beauty_score,
                "beauty_raw_score": round(raw_score, 4),
                "age_model_used": "HowCuteAmI",
                "gender_model_used": "HowCuteAmI",
                "beauty_model_used": "HowCuteAmI",
            }
        except Exception as e:
            logger.error(f"HowCuteAmI beauty gender prediction failed: {e}")
            raise e

    def _predict_age_emotion_with_deepface(
        self, face_image: np.ndarray, include_emotion: bool = True,
        initial_age_confidence: Optional[float] = None
    ) -> Dict[str, Any]:
        """使用DeepFace预测年龄、情绪（并返回可用的性别信息用于回退）"""
        if not DEEPFACE_AVAILABLE:
            # 如果DeepFace不可用，使用HowCuteAmI的年龄预测作为回退
            return self._predict_age_with_howcuteami_fallback(face_image)

        if face_image is None or face_image.size == 0:
            raise ValueError("无效的人脸图像")

        try:
            actions = ["age", "gender", "emotion"] if include_emotion else ["age", "gender"]
            # DeepFace分析 - 禁用进度条和详细输出
            result = DeepFace.analyze(
                img_path=face_image,
                actions=actions,
                enforce_detection=False,
                detector_backend="skip",
                silent=True  # 禁用进度条输出
            )

            # 处理结果 (DeepFace返回的结果格式可能是list或dict)
            if isinstance(result, list):
                result = result[0]

            # 提取信息
            age = result.get("age", 25)
            if include_emotion:
                emotion = result.get("dominant_emotion", "neutral")
                emotion_scores = result.get("emotion", {}) or {}
            else:
                emotion = "neutral"
                emotion_scores = {"neutral": 100.0}
            # 性别信息（用于在HowCuteAmI置信度低时回退）
            deep_gender = result.get("dominant_gender", "Woman")
            deep_gender_conf = result.get("gender", {}).get(deep_gender, 50.0) / 100.0
            deep_gender_conf = self._cap_conf(deep_gender_conf)
            if str(deep_gender).lower() in ["woman", "female"]:
                deep_gender = "Female"
            else:
                deep_gender = "Male"

            # 使用传入的置信度，如果没有则使用固定参考值 0.85
            age_conf = initial_age_confidence if initial_age_confidence is not None else 0.85
            return {
                "age": str(int(age)),
                "age_confidence": round(age_conf, 4),
                "emotion": emotion,
                "emotion_analysis": emotion_scores,
                "gender": deep_gender,
                "gender_confidence": deep_gender_conf,
            }
        except Exception as e:
            logger.error(f"DeepFace age emotion prediction failed, falling back to HowCuteAmI: {e}")
            return self._predict_age_with_howcuteami_fallback(face_image)

    def _predict_age_with_howcuteami_fallback(
        self, face_image: np.ndarray
    ) -> Dict[str, Any]:
        """HowCuteAmI年龄预测回退方案"""
        try:
            if face_image is None or face_image.size == 0:
                raise ValueError("无法读取人脸图像")

            face_resized = cv2.resize(face_image, (224, 224))
            blob = cv2.dnn.blobFromImage(
                face_resized, 1.0, (224, 224), self.MODEL_MEAN_VALUES, swapRB=False
            )

            # 年龄预测
            self.age_net.setInput(blob)
            age_preds = self.age_net.forward()
            age = self.age_list[age_preds[0].argmax()]
            age_confidence = float(np.max(age_preds[0]))

            return {
                "age": age[1:-1],  # 去掉括号
                "age_confidence": round(age_confidence, 4),
                "emotion": "neutral",  # 默认情绪
                "emotion_analysis": {"neutral": 100.0},  # 默认情绪分析
            }
        except Exception as e:
            logger.error(f"HowCuteAmI age prediction fallback failed: {e}")
            return {
                "age": "25-32",
                "age_confidence": 0.5,
                "emotion": "neutral",
                "emotion_analysis": {"neutral": 100.0},
            }

    def _predict_with_hybrid_model(
        self, face: np.ndarray, face_image: np.ndarray
    ) -> Dict[str, Any]:
        """混合模型预测：HowCuteAmI（颜值+性别）+ DeepFace（年龄+情绪，年龄置信度低时优先使用）"""
        # 使用HowCuteAmI预测颜值和性别
        beauty_gender_result = self._predict_beauty_gender_with_howcuteami(face)

        # Hybrid 模式下可配置是否启用 DeepFace 情绪识别（默认启用）。
        deepface_emotion_enabled = bool(getattr(config, "DEEPFACE_EMOTION_ENABLED", True))
        age_emotion_result: Optional[Dict[str, Any]] = None

        # 首先获取HowCuteAmI的年龄/性别预测置信度
        howcuteami_age_confidence = beauty_gender_result.get("age_confidence", 0)
        gender_confidence = beauty_gender_result.get("gender_confidence", 0)
        if gender_confidence >= 1:
            gender_confidence = 0.9999
        age = beauty_gender_result["age"]

        # 如果HowCuteAmI的年龄置信度低于阈值，则使用DeepFace的年龄
        agec = config.AGE_CONFIDENCE
        if howcuteami_age_confidence < agec:
            # 需要 DeepFace 年龄时，仍调用 DeepFace（但可按开关选择是否同时跑 emotion）
            age_emotion_result = self._predict_age_emotion_with_deepface(
                face_image, include_emotion=deepface_emotion_enabled,
                initial_age_confidence=howcuteami_age_confidence
            )
            deep_age = age_emotion_result["age"]
            logger.info(
                f"HowCuteAmI age confidence ({howcuteami_age_confidence}) below {agec}, value=({age}); using DeepFace for age prediction, value={deep_age}"
            )
            # 合并结果，使用DeepFace的年龄预测
            result = {
                "gender": beauty_gender_result["gender"],  # 先用HowCuteAmI，后面可能回退
                "gender_confidence": self._cap_conf(gender_confidence),
                "beauty_score": beauty_gender_result["beauty_score"],
                "beauty_raw_score": beauty_gender_result["beauty_raw_score"],
                "age": deep_age,
                "age_confidence": age_emotion_result["age_confidence"],
                "emotion": age_emotion_result.get("emotion") or "neutral",
                "emotion_analysis": age_emotion_result.get("emotion_analysis") or {"neutral": 100.0},
                "model_used": "hybrid_deepface_age",
                "age_model_used": "DeepFace",
                "gender_model_used": "HowCuteAmI",
            }
        else:
            # HowCuteAmI年龄置信度足够高，使用原有逻辑
            logger.info(
                f"HowCuteAmI age confidence ({howcuteami_age_confidence}) is high enough, value={age}; using HowCuteAmI for age prediction"
            )
            # 情绪识别完全可选：关闭时直接返回默认值，避免多一次 DeepFace 推理。
            if deepface_emotion_enabled:
                age_emotion_result = self._predict_age_emotion_with_deepface(
                    face_image, include_emotion=True,
                    initial_age_confidence=howcuteami_age_confidence
                )
                emotion = age_emotion_result.get("emotion") or "neutral"
                emotion_analysis = age_emotion_result.get("emotion_analysis") or {"neutral": 100.0}
            else:
                emotion = "neutral"
                emotion_analysis = {"neutral": 100.0}
            # 合并结果，保留HowCuteAmI的年龄预测
            result = {
                "gender": beauty_gender_result["gender"],  # 先用HowCuteAmI，后面可能回退
                "gender_confidence": self._cap_conf(gender_confidence),
                "beauty_score": beauty_gender_result["beauty_score"],
                "beauty_raw_score": beauty_gender_result["beauty_raw_score"],
                "age": beauty_gender_result["age"],
                "age_confidence": beauty_gender_result["age_confidence"],
                "emotion": emotion,
                "emotion_analysis": emotion_analysis,
                "model_used": "hybrid",
                "age_model_used": "HowCuteAmI",
                "gender_model_used": "HowCuteAmI",
            }

        # 统一性别判定规则：任一模型判为Female则Female；两者都为Male才Male
        try:
            how_gender = beauty_gender_result.get("gender")
            how_conf = float(beauty_gender_result.get("gender_confidence", 0) or 0)
            deep_gender = age_emotion_result.get("gender") if age_emotion_result else None
            deep_conf = float(age_emotion_result.get("gender_confidence", 0) or 0) if age_emotion_result else 0.0

            final_gender = result.get("gender")
            final_conf = float(result.get("gender_confidence", 0) or 0)
            # 规则判断
            if (str(how_gender) == "Female") or (str(deep_gender) == "Female"):
                final_gender = "Female"
                final_conf = max(how_conf if how_gender == "Female" else 0,
                                 deep_conf if deep_gender == "Female" else 0)
                result["gender_model_used"] = "Combined(H+DF)"
            elif (str(how_gender) == "Male") and (str(deep_gender) == "Male"):
                final_gender = "Male"
                final_conf = max(how_conf if how_gender == "Male" else 0,
                                 deep_conf if deep_gender == "Male" else 0)
                result["gender_model_used"] = "Combined(H+DF)"
            # 否则保持原判定

            result["gender"] = final_gender
            result["gender_confidence"] = self._cap_conf(final_conf)
        except Exception:
            pass

        return result

    def _predict_with_howcuteami(self, face: np.ndarray) -> Dict[str, Any]:
        """使用HowCuteAmI模型进行完整预测"""
        try:
            # 性别预测
            blob = cv2.dnn.blobFromImage(
                face, 1.0, (224, 224), self.MODEL_MEAN_VALUES, swapRB=False
            )
            self.gender_net.setInput(blob)
            gender_preds = self.gender_net.forward()
            gender = self.gender_list[gender_preds[0].argmax()]
            gender_confidence = float(np.max(gender_preds[0]))
            gender_confidence = self._cap_conf(gender_confidence)

            # 年龄预测
            self.age_net.setInput(blob)
            age_preds = self.age_net.forward()
            age = self.age_list[age_preds[0].argmax()]
            age_confidence = float(np.max(age_preds[0]))

            # 颜值预测
            blob_beauty = cv2.dnn.blobFromImage(
                face, 1.0 / 255, (224, 224), self.MODEL_MEAN_VALUES, swapRB=False
            )
            self.beauty_net.setInput(blob_beauty)
            beauty_preds = self.beauty_net.forward()
            beauty_score = round(float(2.0 * np.sum(beauty_preds[0])), 1)
            beauty_score = min(10.0, max(0.0, beauty_score))
            beauty_score = self._adjust_beauty_score(beauty_score)
            raw_score = float(np.sum(beauty_preds[0]))

            return {
                "gender": gender,
                "gender_confidence": gender_confidence,
                "age": age[1:-1],  # 去掉括号
                "age_confidence": round(age_confidence, 4),
                "beauty_score": beauty_score,
                "beauty_raw_score": round(raw_score, 4),
                "model_used": "HowCuteAmI",
                "emotion": "neutral",  # HowCuteAmI不支持情绪分析
                "emotion_analysis": {"neutral": 100.0},
                "age_model_used": "HowCuteAmI",
                "gender_model_used": "HowCuteAmI",
                "beauty_model_used": "HowCuteAmI",
            }
        except Exception as e:
            logger.error(f"HowCuteAmI prediction failed: {e}")
            raise e

    def _predict_with_deepface(self, face_image: np.ndarray) -> Dict[str, Any]:
        """使用DeepFace进行预测"""
        if not DEEPFACE_AVAILABLE:
            raise ValueError("DeepFace未安装")

        if face_image is None or face_image.size == 0:
            raise ValueError("无效的人脸图像")

        try:
            deepface_emotion_enabled = bool(getattr(config, "DEEPFACE_EMOTION_ENABLED", True))
            actions = ["age", "gender", "emotion"] if deepface_emotion_enabled else ["age", "gender"]
            # DeepFace分析 - 禁用进度条和详细输出
            result = DeepFace.analyze(
                img_path=face_image,
                actions=actions,
                enforce_detection=False,
                detector_backend="skip",
                silent=True  # 禁用进度条输出
            )

            # 处理结果 (DeepFace返回的结果格式可能是list或dict)
            if isinstance(result, list):
                result = result[0]

            # 提取信息
            age = result.get("age", 25)
            gender = result.get("dominant_gender", "Woman")
            gender_confidence = result.get("gender", {}).get(gender, 0.5) / 100
            gender_confidence = self._cap_conf(gender_confidence)

            # 统一性别标签
            if gender.lower() in ["woman", "female"]:
                gender = "Female"
            else:
                gender = "Male"

            # DeepFace没有内置颜值评分，这里使用简单的启发式方法
            if deepface_emotion_enabled:
                emotion = result.get("dominant_emotion", "neutral")
                emotion_scores = result.get("emotion", {}) or {}
            else:
                emotion = "neutral"
                emotion_scores = {"neutral": 100.0}

            # 基于情绪和年龄的简单颜值估算
            happiness_score = emotion_scores.get("happy", 0) / 100
            neutral_score = emotion_scores.get("neutral", 0) / 100

            # 简单的颜值算法 (可以改进)
            base_beauty = 6.0  # 基础分
            emotion_bonus = happiness_score * 2 + neutral_score * 1
            age_factor = max(0.5, 1 - abs(age - 25) / 50)  # 25岁为最佳年龄

            beauty_score = round(min(10.0, base_beauty + emotion_bonus + age_factor), 2)

            # DeepFace不返回年龄置信度，使用固定参考值 0.85
            age_conf = 0.85
            return {
                "gender": gender,
                "gender_confidence": gender_confidence,
                "age": str(int(age)),
                "age_confidence": age_conf,
                "beauty_score": beauty_score,
                "beauty_raw_score": round(beauty_score / 10, 4),
                "model_used": "DeepFace",
                "emotion": emotion,
                "emotion_analysis": emotion_scores,
                "age_model_used": "DeepFace",
                "gender_model_used": "DeepFace",
                "beauty_model_used": "Heuristic",
            }
        except Exception as e:
            logger.error(f"DeepFace prediction failed: {e}")
            raise e

    def analyze_faces(
        self,
        image: np.ndarray,
        original_image_hash: str,
        model_type: ModelType = ModelType.HYBRID,
    ) -> Dict[str, Any]:
        """
        分析图片中的人脸
        :param image: 输入图像
        :param original_image_hash: 原始图片的MD5哈希值
        :param model_type: 使用的模型类型
        :return: 分析结果
        """
        if image is None:
            raise ValueError("invalid image input")

        # 检测人脸
        face_boxes = self._detect_faces(image)

        if not face_boxes:
            return {
                "success": False,
                "message": "Please upload a clear front-facing photo without any obstructions.",
                "face_count": 0,
                "faces": [],
                "annotated_image": None,
                "model_used": model_type.value,
            }

        results = {
            "success": True,
            "message": f"success detected {len(face_boxes)} faces",
            "face_count": len(face_boxes),
            "faces": [],
            "model_used": model_type.value,
        }

        # 复制原图用于绘制
        annotated_image = image.copy()
        logger.info(
            f"Input annotated_image shape: {annotated_image.shape}, dtype: {annotated_image.dtype}, ndim: {annotated_image.ndim}"
        )
        # 分析每张人脸
        for i, face_box in enumerate(face_boxes):
            # 裁剪人脸
            face_cropped = self._crop_face(image, face_box)
            if face_cropped.size == 0:
                logger.warning(f"Cropped face {i + 1} is empty, skipping.")
                continue

            face_resized = cv2.resize(face_cropped, (224, 224))
            face_for_deepface = face_cropped.copy()

            # 根据模型类型进行预测
            try:
                if model_type == ModelType.HYBRID:
                    # 混合模式：颜值性别用HowCuteAmI，年龄情绪用DeepFace
                    prediction_result = self._predict_with_hybrid_model(
                        face_resized, face_for_deepface
                    )
                elif model_type == ModelType.HOWCUTEAMI:
                    prediction_result = self._predict_with_howcuteami(face_resized)
                    # 非混合模式也进行性别合并：引入DeepFace性别（不需要 emotion，减少耗时）
                    try:
                        age_emotion_result = self._predict_age_emotion_with_deepface(
                            face_for_deepface, include_emotion=False
                        )
                        how_gender = prediction_result.get("gender")
                        how_conf = float(prediction_result.get("gender_confidence", 0) or 0)
                        deep_gender = age_emotion_result.get("gender")
                        deep_conf = float(age_emotion_result.get("gender_confidence", 0) or 0)
                        final_gender = prediction_result.get("gender")
                        final_conf = float(prediction_result.get("gender_confidence", 0) or 0)
                        if (str(how_gender) == "Female") or (str(deep_gender) == "Female"):
                            final_gender = "Female"
                            final_conf = max(
                                how_conf if how_gender == "Female" else 0,
                                deep_conf if deep_gender == "Female" else 0,
                            )
                            prediction_result["gender_model_used"] = "Combined(H+DF)"
                        elif (str(how_gender) == "Male") and (str(deep_gender) == "Male"):
                            final_gender = "Male"
                            final_conf = max(
                                how_conf if how_gender == "Male" else 0,
                                deep_conf if deep_gender == "Male" else 0,
                            )
                            prediction_result["gender_model_used"] = "Combined(H+DF)"
                        prediction_result["gender"] = final_gender
                        prediction_result["gender_confidence"] = round(float(final_conf), 4)
                    except Exception:
                        pass
                elif model_type == ModelType.DEEPFACE and DEEPFACE_AVAILABLE:
                    prediction_result = self._predict_with_deepface(face_for_deepface)
                    # 非混合模式也进行性别合并：引入HowCuteAmI性别
                    try:
                        beauty_gender_result = self._predict_beauty_gender_with_howcuteami(
                            face_resized
                        )
                        deep_gender = prediction_result.get("gender")
                        deep_conf = float(prediction_result.get("gender_confidence", 0) or 0)
                        how_gender = beauty_gender_result.get("gender")
                        how_conf = float(beauty_gender_result.get("gender_confidence", 0) or 0)
                        final_gender = prediction_result.get("gender")
                        final_conf = float(prediction_result.get("gender_confidence", 0) or 0)
                        if (str(how_gender) == "Female") or (str(deep_gender) == "Female"):
                            final_gender = "Female"
                            final_conf = max(
                                how_conf if how_gender == "Female" else 0,
                                deep_conf if deep_gender == "Female" else 0,
                            )
                            prediction_result["gender_model_used"] = "Combined(H+DF)"
                        elif (str(how_gender) == "Male") and (str(deep_gender) == "Male"):
                            final_gender = "Male"
                            final_conf = max(
                                how_conf if how_gender == "Male" else 0,
                                deep_conf if deep_gender == "Male" else 0,
                            )
                            prediction_result["gender_model_used"] = "Combined(H+DF)"
                        prediction_result["gender"] = final_gender
                        prediction_result["gender_confidence"] = round(float(final_conf), 4)
                    except Exception:
                        pass
                else:
                    # 回退到混合模式
                    prediction_result = self._predict_with_hybrid_model(
                        face_resized, face_for_deepface
                    )
                    logger.warning(
                        f"Model {model_type.value} is not available, using hybrid mode"
                    )

            except Exception as e:
                logger.error(f"Prediction failed, using default values: {e}")
                prediction_result = {
                    "gender": "Unknown",
                    "gender_confidence": 0.5,
                    "age": "25-32",
                    "age_confidence": 0.5,
                    "beauty_score": 5.0,
                    "beauty_raw_score": 0.5,
                    "emotion": "neutral",
                    "emotion_analysis": {"neutral": 100.0},
                    "model_used": "fallback",
                }

            # 五官分析
            # facial_features = self.facial_analyzer.analyze_facial_features(
            #     face_cropped, face_box
            # )

            # 颜色设置与年龄显示统一（应用女性年龄调整）
            gender = prediction_result.get("gender", "Unknown")
            color_bgr = self.gender_colors.get(gender, (128, 128, 128))
            color_hex = f"#{color_bgr[2]:02x}{color_bgr[1]:02x}{color_bgr[0]:02x}"

            # 年龄文本与调整
            raw_age_str = prediction_result.get("age", "Unknown")
            display_age_str = str(raw_age_str)
            age_adjusted_flag = False
            age_adjustment_value = int(getattr(config, "FEMALE_AGE_ADJUSTMENT", 0) or 0)
            age_adjustment_threshold = int(getattr(config, "FEMALE_AGE_ADJUSTMENT_THRESHOLD", 999) or 999)

            # 仅对女性且年龄达到阈值时进行调整
            try:
                # 支持 "25-32" 或 "25" 格式
                if "-" in str(raw_age_str):
                    age_num = int(str(raw_age_str).split("-")[0].strip("() "))
                else:
                    age_num = int(str(raw_age_str).strip())

                if str(gender) == "Female" and age_num >= age_adjustment_threshold and age_adjustment_value > 0:
                    adjusted_age = max(0, age_num - age_adjustment_value)
                    display_age_str = str(adjusted_age)
                    age_adjusted_flag = True
                    try:
                        logger.info(f"Adjusted age for female (draw+data): {age_num} -> {adjusted_age}")
                    except Exception:
                        pass
            except Exception:
                # 无法解析年龄时，保持原样
                pass

            # 保存裁剪的人脸
            date_sub = get_date_subfolder(OUTPUT_DIR)
            cropped_face_filename = f"{date_sub}/{original_image_hash}_face_{i + 1}.webp"
            cropped_face_path = os.path.join(OUTPUT_DIR, cropped_face_filename)
            try:
                save_image_high_quality(face_cropped, cropped_face_path)
                logger.info(f"cropped face: {cropped_face_path}")
            except Exception as e:
                logger.error(f"Failed to save cropped face {cropped_face_path}: {e}")
                cropped_face_filename = None

            # 在图片上绘制标注
            if config.DRAW_SCORE:
                cv2.rectangle(
                    annotated_image,
                    (face_box[0], face_box[1]),
                    (face_box[2], face_box[3]),
                    color_bgr,
                    int(round(image.shape[0] / 400)),
                    8,
                )

            # 标签文本
            beauty_score = prediction_result.get("beauty_score", 0)
            label = f"{gender}, {display_age_str}, {beauty_score}"

            font_scale = max(
                0.3, min(0.7, image.shape[0] / 800)
            )  # 从500改为800，范围从0.5-1.0改为0.3-0.7
            font_thickness = 2
            font = cv2.FONT_HERSHEY_SIMPLEX
            # 绘制文本
            text_x = face_box[0]
            text_y = face_box[1] - 10 if face_box[1] - 10 > 20 else face_box[1] + 30

            # 计算文字大小（宽高）
            (text_width, text_height), baseline = cv2.getTextSize(label, font, font_scale, font_thickness)

            # 画黑色矩形背景，稍微比文字框大一点，增加边距
            background_tl = (text_x, text_y - text_height - baseline)  # 矩形左上角
            background_br = (text_x + text_width, text_y + baseline)  # 矩形右下角

            if config.DRAW_SCORE:
                cv2.rectangle(
                    annotated_image,
                    background_tl,
                    background_br,
                    color_bgr,  # 黑色背景
                    thickness=-1  # 填充
                )
                cv2.putText(
                    annotated_image,
                    label,
                    (text_x, text_y),
                    font,
                    font_scale,
                    (255, 255, 255),
                    font_thickness,
                    cv2.LINE_AA,
                )

            # 构建人脸结果
            face_result = {
                "face_id": i + 1,
                "gender": gender,
                "gender_confidence": prediction_result.get("gender_confidence", 0),
                "gender_model_used": prediction_result.get("gender_model_used", prediction_result.get("model_used", model_type.value)),
                "age": display_age_str,
                "age_confidence": prediction_result.get("age_confidence", 0),
                "age_model_used": prediction_result.get("age_model_used", prediction_result.get("model_used", model_type.value)),
                "beauty_score": prediction_result.get("beauty_score", 0),
                "beauty_raw_score": prediction_result.get("beauty_raw_score", 0),
                "emotion": prediction_result.get("emotion") or "neutral",
                "emotion_analysis": prediction_result.get("emotion_analysis") or {"neutral": 100.0},
                # "facial_features": facial_features,  # 五官分析
                "bounding_box": {
                    "x1": int(face_box[0]),
                    "y1": int(face_box[1]),
                    "x2": int(face_box[2]),
                    "y2": int(face_box[3]),
                },
                "color": {
                    "bgr": [int(color_bgr[0]), int(color_bgr[1]), int(color_bgr[2])],
                    "hex": color_hex,
                },
                "cropped_face_filename": cropped_face_filename,
                "model_used": prediction_result.get("model_used", model_type.value),
            }

            if age_adjusted_flag:
                face_result["age_adjusted"] = True
                face_result["age_adjustment_value"] = int(age_adjustment_value)

            results["faces"].append(face_result)

        results["annotated_image"] = annotated_image
        return results

    def _warmup_models(self):
        """预热模型，减少首次调用延迟"""
        try:
            logger.info("Starting to warm up models...")

            # 创建一个小的测试图像 (64x64)
            test_image = np.ones((64, 64, 3), dtype=np.uint8) * 128

            # 预热DeepFace模型（如果可用）
            if DEEPFACE_AVAILABLE:
                try:
                    import tempfile
                    deepface_emotion_enabled = bool(getattr(config, "DEEPFACE_EMOTION_ENABLED", True))
                    with tempfile.NamedTemporaryFile(suffix='.webp', delete=False) as tmp_file:
                        cv2.imwrite(tmp_file.name, test_image, [cv2.IMWRITE_WEBP_QUALITY, 95])
                        # 预热DeepFace - 使用最小的actions集合
                        DeepFace.analyze(
                            img_path=tmp_file.name,
                            actions=["age", "gender", "emotion"] if deepface_emotion_enabled else ["age", "gender"],
                            detector_backend="yolov8",
                            enforce_detection=False,
                            silent=True
                        )
                        os.unlink(tmp_file.name)
                    logger.info("DeepFace model warm-up completed")
                except Exception as e:
                    logger.warning(f"DeepFace model warm-up failed: {e}")

            # 预热OpenCV DNN模型
            try:
                # 预热人脸检测模型
                blob = cv2.dnn.blobFromImage(test_image, 1.0, (300, 300), (104, 117, 123))
                self.face_net.setInput(blob)
                self.face_net.forward()

                # 预热年龄预测模型
                test_face = cv2.resize(test_image, (224, 224))
                blob = cv2.dnn.blobFromImage(test_face, 1.0, (224, 224), self.MODEL_MEAN_VALUES, swapRB=False)
                self.age_net.setInput(blob)
                self.age_net.forward()

                # 预热性别预测模型
                self.gender_net.setInput(blob)
                self.gender_net.forward()

                # 预热颜值评分模型
                self.beauty_net.setInput(blob)
                self.beauty_net.forward()

                logger.info("OpenCV DNN model warm-up completed")
            except Exception as e:
                logger.warning(f"OpenCV DNN model warm-up failed: {e}")

            logger.info("Model warm-up completed")
        except Exception as e:
            logger.warning(f"Error occurred during model warm-up: {e}")