Captured at (local ISO): 2026-05-18 05:16:54

I. Preface

Face-attribute analysis—gender, age, and more—is a core CV topic for surveillance, user profiling, retail footfall analytics, etc. Classical pipelines often fail on wild faces: weak detectors, heavy deploy friction.

I built this while scaling a dataset: after face clustering to dedupe speakers, the next step was gender/age distribution stats. This note documents that system end-to-end with source.

The stack uses MTCNN (Multi-task Cascaded CNN) for robust detection plus off-the-shelf Caffe nets for gender and 8-bin age estimation, then writes a structured text report. It batch-processes every image under a folder, tallying face count, gender histogram, age-bin histogram, mean age (mid-bin proxy), and paths with multiple faces. Useful for learning CV or production-style footfall analytics.

II. Model overview

Two stages: detect, then classify.

2.1 MTCNN: high-accuracy face detection

MTCNN (Multi-task Cascaded Convolutional Neural Networks, 2016) jointly optimizes face detection and landmark localization—fast in the wild.

• Highlights
  1. Cascade: P-Net → R-Net → O-Net progressively refine boxes—fewer FPs/FNs.
  2. Multi-task: predicts 5 (or 68) landmarks (eyes, nose, mouth corners)—handy for alignment (unused here but easy to add).
  3. Light: small weights, good throughput for batch jobs.
• In this repo: replaces a vanilla OpenCV DNN face detector for profile, occlusion, tiny faces; default confidence cutoff 0.6 (tune as needed).

2.2 Caffe pretrained models: gender & age

Two separate Caffe nets—classic CNN heads trained on public corpora (e.g., IMDB–WIKI), decent out-of-domain behavior.

• Bins → mid values for mean-age math:
  • (0-2)→1; (4-6)→5; (8-12)→10; (15-20)→18; (25-32)→28; (38-43)→40; (48-53)→50; (60-100)→70

III. How the models work

3.1 MTCNN pipeline

Three stages, each refining the last:

1. P-Net
   • In: image pyramid.
   • Job: cheap candidate boxes + face scores + box regressions.
   • Out: NMS-filtered candidates.
2. R-Net
   • In: P-Net crops resized to 24×24.
   • Job: reject bad proposals, refine boxes.
   • Out: fewer, tighter boxes.
3. O-Net
   • In: R-Net crops to 48×48.
   • Job: final box regression + 5 landmarks.
   • Out: final boxes (we consume boxes only; landmarks are free for later).

Coarse → fine → final keeps speed and robustness in clutter.

3.2 Caffe attribute heads

Same recipe for gender and age:

1. Preprocess
   • Crop MTCNN box → 227×227.
   • Subtract per-channel means (78.426337, 87.768914, 114.895847) in BGR.
   • Pack a blob N×C×H×W.
2. Features
   • Convs / ReLU / pools build hierarchical cues—edges/textures low level; parts high level.
3. Heads
   • Gender: 2-way softmax → argmax label.
   • Age: 8-way softmax → bin label → mid value for averages.

IV. Code walkthrough

4.1 Overall layout

Modules: paths, loaders, core functions, main.

4.2 Paths & imports

import os
import cv2
import numpy as np
from collections import Counter, defaultdict
from datetime import datetime
from mtcnn import MTCNN   # 人脸检测库
from tqdm import tqdm  # 进度条库

# -------- 路径配置 --------
THIS_DIR = os.path.dirname(__file__)  # 当前脚本所在目录
PROJECT_ROOT = os.path.abspath(os.path.join(THIS_DIR, '..'))  # 项目根目录
AGE_GENDER_DIR = os.path.join(PROJECT_ROOT, 'age_gender')  # 性别/年龄模型目录
FACE_DIR = os.path.join(PROJECT_ROOT, 'face', 'face_dir')  # 待处理图片目录
RESULT_TXT = os.path.join(PROJECT_ROOT, 'face', '性别及年龄统计.txt')  # 结果输出文件

# -------- 模型文件路径 --------
AGE_PROTO = os.path.join(AGE_GENDER_DIR, 'age_deploy.prototxt')  # 年龄模型配置文件
AGE_MODEL = os.path.join(AGE_GENDER_DIR, 'age_net.caffemodel')  # 年龄模型权重
GENDER_PROTO = os.path.join(AGE_GENDER_DIR, 'gender_deploy.prototxt')  # 性别模型配置文件
GENDER_MODEL = os.path.join(AGE_GENDER_DIR, 'gender_net.caffemodel')  # 性别模型权重

# 年龄分段与中间值配置
AGE_LIST = ['(0-2)', '(4-6)', '(8-12)', '(15-20)', '(25-32)', '(38-43)', '(48-53)', '(60-100)']
AGE_MID = [1, 5, 10, 18, 28, 40, 50, 70]  # 用于计算平均年龄
GENDER_LIST = ['Female', 'Male']  # 性别选项

# 全局MTCNN检测器（避免重复初始化，提升效率）
mtcnn_detector = MTCNN()

• Role: wire directories, Caffe artifacts, age metadata, singleton MTCNN (avoid per-image init).
• Deps: pip install mtcnn opencv-python tqdm.

4.3 load_models

def load_models():
    # 仅加载年龄/性别模型；人脸检测改为全局MTCNN
    age_net = cv2.dnn.readNet(AGE_MODEL, AGE_PROTO)
    gender_net = cv2.dnn.readNet(GENDER_MODEL, GENDER_PROTO)
    return None, age_net, gender_net  # 第一个位置返回None，兼容旧调用逻辑

• Loads age/gender via cv2.dnn.readNet.
• Leading None preserves older tuple unpacking when OpenCV supplied a face net—now unused.

4.4 detect_faces

def detect_faces(face_net, img, conf_thresh=0.6):
    """
    使用MTCNN进行人脸检测
    参数:
        face_net: 兼容旧参数（现未使用）
        img: 输入图像（BGR格式，OpenCV默认读取格式）
        conf_thresh: 检测置信度阈值
    返回:
        boxes: 人脸框列表，格式为[(x1,y1,x2,y2), ...]（左上角(x1,y1)，右下角(x2,y2)）
    """
    results = mtcnn_detector.detect_faces(img)  # MTCNN检测结果
    boxes = []
    h, w = img.shape[:2]  # 图像高度和宽度
    for r in results:
        confidence = r.get('confidence', 0)  # 检测置信度
        if confidence >= conf_thresh and 'box' in r:
            x, y, w_box, h_box = r['box']  # MTCNN返回的框：x,y为左上角坐标，w_box/h_box为宽高
            # 修正框坐标（避免超出图像边界）
            x1 = max(0, x)
            y1 = max(0, y)
            x2 = min(x + w_box, w - 1)
            y2 = min(y + h_box, h - 1)
            # 确保框为有效区域（宽高均大于0）
            if x2 > x1 and y2 > y1:
                boxes.append((x1, y1, x2, y2))
    return boxes

• Filters by confidence, clamps to frame, drops degenerate boxes.
• face_net remains for backward-compatible calls.

4.5 predict_age_gender

def predict_age_gender(age_net, gender_net, face_img):
    """
    预测单张人脸的性别和年龄
    参数:
        age_net: 年龄预测模型
        gender_net: 性别预测模型
        face_img: 裁剪后的人脸图像（BGR格式）
    返回:
        gender: 预测性别（Female/Male）
        age: 预测年龄段（如(25-32)）
        age_mid: 年龄段中间值（如28）
    """
    # 图像预处理：转换为Blob格式并减去均值
    blob = cv2.dnn.blobFromImage(
        face_img, 
        1.0,  # 缩放因子
        (227, 227),  # 模型输入尺寸
        (78.426337, 87.768914, 114.895847),  # 各通道均值（BGR顺序）
        swapRB=False  # 无需交换R和B通道（OpenCV读取为BGR，模型要求BGR）
    )
    
    # 性别预测
    gender_net.setInput(blob)  # 设置模型输入
    gender_pred = gender_net.forward()  # 推理得到预测结果
    gender = GENDER_LIST[gender_pred[0].argmax()]  # 取概率最大的类别
    
    # 年龄预测
    age_net.setInput(blob)  # 设置模型输入
    age_pred = age_net.forward()  # 推理得到预测结果
    age_idx = age_pred[0].argmax()  # 取概率最大的年龄段索引
    age = AGE_LIST[age_idx]  # 年龄段标签
    age_mid = AGE_MID[age_idx]  # 年龄段中间值
    
    return gender, age, age_mid

Steps: blobFromImage → forward twice → argmax maps.

4.6 iter_images

def iter_images(root):
    """
    遍历目录下所有支持的图片文件（递归遍历子目录）
    参数:
        root: 根目录路径
    返回:
        图片文件路径生成器
    """
    exts = {'.jpg', '.jpeg', '.png', '.bmp'}  # 支持的图片格式
    for dirpath, _, filenames in os.walk(root):  # 递归遍历目录
        for f in filenames:
            # 检查文件后缀是否在支持的格式中（不区分大小写）
            if os.path.splitext(f.lower())[1] in exts:
                yield os.path.join(dirpath, f)  # 返回完整文件路径

Recursive generator—low memory.

4.7 main

Orchestrates enumerate → detect → infer → aggregate → save.

def main():
    if not os.path.exists(FACE_DIR):
        print("face_dir 不存在:", FACE_DIR)
        return
    face_net, age_net, gender_net = load_models()  # face_net 现在恒为 None
    gender_counter = Counter()
    age_counter = Counter()
    age_mid_values = []
    file_face_count = defaultdict(int)
    total_faces = 0

    # 预收集所有图片路径用于进度条
    image_paths = list(iter_images(FACE_DIR))
    total_images = len(image_paths)
    if total_images == 0:
        print("未找到任何图片，目录:", FACE_DIR)
        return

    for img_path in tqdm(image_paths, desc='处理图片', unit='张'):
        img = cv2.imread(img_path)
        if img is None:
            continue
        boxes = detect_faces(face_net, img)
        for (x1, y1, x2, y2) in boxes:
            face = img[y1:y2, x1:x2]
            if face.size == 0:
                continue
            try:
                gender, age_range, age_mid = predict_age_gender(age_net, gender_net, face)
            except Exception as e:
                # 单张失败不影响整体
                continue
            gender_counter[gender] += 1
            age_counter[age_range] += 1
            age_mid_values.append(age_mid)
            file_face_count[img_path] += 1
            total_faces += 1

    avg_age = round(sum(age_mid_values) / len(age_mid_values), 2) if age_mid_values else 0.0

    lines = []
    lines.append("统计时间: " + datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
    lines.append(f"扫描图片数: {total_images}")
    lines.append(f"检测到人脸总数: {total_faces}")
    lines.append("性别统计:")
    for g in GENDER_LIST:
        lines.append(f"  {g}: {gender_counter.get(g,0)}")
    lines.append("年龄段统计:")
    for a in AGE_LIST:
        lines.append(f"  {a}: {age_counter.get(a,0)}")
    lines.append(f"估算平均年龄: {avg_age}")
    lines.append("（文件级人脸计数仅列出>1者）")
    for fp, c in file_face_count.items():
        if c > 1:
            lines.append(f"  {fp}: {c}")

    os.makedirs(os.path.dirname(RESULT_TXT), exist_ok=True)
    with open(RESULT_TXT, 'w', encoding='utf-8') as f:
        f.write('\n'.join(lines))

    print("统计完成，结果写入:", RESULT_TXT)
    print('\n'.join(lines[:10]), '...')

4.8 Full listing

import os
import cv2
import numpy as np
from collections import Counter, defaultdict
from datetime import datetime
from mtcnn import MTCNN   # 新增：使用与 face_check.py 相同的 MTCNN 检测
from tqdm import tqdm  # 进度条

# -------- 路径配置 --------
THIS_DIR = os.path.dirname(__file__)
PROJECT_ROOT = os.path.abspath(os.path.join(THIS_DIR, '..'))
AGE_GENDER_DIR = os.path.join(PROJECT_ROOT, 'age_gender')
FACE_DIR = os.path.join(PROJECT_ROOT, 'face', 'face_dir')
RESULT_TXT = os.path.join(PROJECT_ROOT, 'face', '性别及年龄统计.txt')

# -------- 模型文件 --------
AGE_PROTO = os.path.join(AGE_GENDER_DIR, 'age_deploy.prototxt')
AGE_MODEL = os.path.join(AGE_GENDER_DIR, 'age_net.caffemodel')
GENDER_PROTO = os.path.join(AGE_GENDER_DIR, 'gender_deploy.prototxt')
GENDER_MODEL = os.path.join(AGE_GENDER_DIR, 'gender_net.caffemodel')
# 下面两个(OpenCV DNN人脸检测)已不再使用，保留不报错；如需可删除
# FACE_PROTO = os.path.join(AGE_GENDER_DIR, 'opencv_face_detector.pbtxt')
# FACE_MODEL = os.path.join(AGE_GENDER_DIR, 'opencv_face_detector_uint8.pb')

AGE_LIST = ['(0-2)', '(4-6)', '(8-12)', '(15-20)', '(25-32)', '(38-43)', '(48-53)', '(60-100)']
AGE_MID = [1, 5, 10, 18, 28, 40, 50, 70]  # 用于估算平均年龄
GENDER_LIST = ['Female', 'Male']
# 新增：全局 MTCNN 检测器，与 face_check.py 保持一致
mtcnn_detector = MTCNN()

def load_models():
    # 仅加载年龄/性别模型；人脸检测改为全局 MTCNN
    age_net = cv2.dnn.readNet(AGE_MODEL, AGE_PROTO)
    gender_net = cv2.dnn.readNet(GENDER_MODEL, GENDER_PROTO)
    return None, age_net, gender_net  # 第一个位置兼容旧调用

def detect_faces(face_net, img, conf_thresh=0.6):
    """
    使用与 face_check.py 相同的 MTCNN 进行检测
    返回: [(x1,y1,x2,y2), ...]
    """
    results = mtcnn_detector.detect_faces(img)
    boxes = []
    h, w = img.shape[:2]
    for r in results:
        confidence = r.get('confidence', 0)
        if confidence >= conf_thresh and 'box' in r:
            x, y, w_box, h_box = r['box']
            x1 = max(0, x)
            y1 = max(0, y)
            x2 = min(x + w_box, w - 1)
            y2 = min(y + h_box, h - 1)
            if x2 > x1 and y2 > y1:
                boxes.append((x1, y1, x2, y2))
    return boxes

def predict_age_gender(age_net, gender_net, face_img):
    blob = cv2.dnn.blobFromImage(face_img, 1.0, (227, 227), (78.426337, 87.768914, 114.895847), swapRB=False)
    gender_net.setInput(blob)
    gender_pred = gender_net.forward()
    gender = GENDER_LIST[gender_pred[0].argmax()]
    age_net.setInput(blob)
    age_pred = age_net.forward()
    age_idx = age_pred[0].argmax()
    age = AGE_LIST[age_idx]
    age_mid = AGE_MID[age_idx]
    return gender, age, age_mid

def iter_images(root):
    exts = {'.jpg', '.jpeg', '.png', '.bmp'}
    for dirpath, _, filenames in os.walk(root):
        for f in filenames:
            if os.path.splitext(f.lower())[1] in exts:
                yield os.path.join(dirpath, f)

def main():
    if not os.path.exists(FACE_DIR):
        print("face_dir 不存在:", FACE_DIR)
        return
    face_net, age_net, gender_net = load_models()  # face_net 现在恒为 None
    gender_counter = Counter()
    age_counter = Counter()
    age_mid_values = []
    file_face_count = defaultdict(int)
    total_faces = 0

    # 预收集所有图片路径用于进度条
    image_paths = list(iter_images(FACE_DIR))
    total_images = len(image_paths)
    if total_images == 0:
        print("未找到任何图片，目录:", FACE_DIR)
        return

    for img_path in tqdm(image_paths, desc='处理图片', unit='张'):
        img = cv2.imread(img_path)
        if img is None:
            continue
        boxes = detect_faces(face_net, img)
        for (x1, y1, x2, y2) in boxes:
            face = img[y1:y2, x1:x2]
            if face.size == 0:
                continue
            try:
                gender, age_range, age_mid = predict_age_gender(age_net, gender_net, face)
            except Exception as e:
                # 单张失败不影响整体
                continue
            gender_counter[gender] += 1
            age_counter[age_range] += 1
            age_mid_values.append(age_mid)
            file_face_count[img_path] += 1
            total_faces += 1

    avg_age = round(sum(age_mid_values) / len(age_mid_values), 2) if age_mid_values else 0.0

    lines = []
    lines.append("统计时间: " + datetime.now().strftime('%Y-%m-%d %H:%M:%S'))
    lines.append(f"扫描图片数: {total_images}")
    lines.append(f"检测到人脸总数: {total_faces}")
    lines.append("性别统计:")
    for g in GENDER_LIST:
        lines.append(f"  {g}: {gender_counter.get(g,0)}")
    lines.append("年龄段统计:")
    for a in AGE_LIST:
        lines.append(f"  {a}: {age_counter.get(a,0)}")
    lines.append(f"估算平均年龄: {avg_age}")
    lines.append("（文件级人脸计数仅列出>1者）")
    for fp, c in file_face_count.items():
        if c > 1:
            lines.append(f"  {fp}: {c}")

    os.makedirs(os.path.dirname(RESULT_TXT), exist_ok=True)
    with open(RESULT_TXT, 'w', encoding='utf-8') as f:
        f.write('\n'.join(lines))

    print("统计完成，结果写入:", RESULT_TXT)
    print('\n'.join(lines[:10]), '...')

if __name__ == "__main__":
    main()

V. Summary

This pipeline targets face-attribute analytics with a detector + classifier split: MTCNN for robust proposals (score gating + box clamping) and Caffe nets for gender plus 8 age bins (mid values for mean age). Batch-scan a directory, emit totals, histograms, mean age, multi-face paths, and a UTF-8 report file. Modular layout—paths, loaders, helpers, main—is easy to read for study or to extend for footfall / demographic style deployments.