dlib检测视频中的人脸并裁剪为图片保存

环境要求

找个带有基本cv配置的虚拟环境安装上dlib依赖的人脸检测的基础环境即可，主要是：

pip install boost dlib opencv-python

缺的按提示安装。

demo

设置好视频路径和图像保存路径，裁剪尺寸（默认256）以及裁剪帧数（默认64），可以直接运行：

import os
import random
import cv2
import dlib
from imutils.face_utils import FaceAligner, rect_to_bb
from tqdm import tqdm  # 引入tqdm库

# 配置路径
dataset_path = r'D:\python_project\face-parsing\dataset'  # 原始数据集路径
output_path = r'D:\python_project\face-parsing\dataset\results'  # 输出路径
crop_size = 256  # 人脸裁剪后的大小


# 获取人脸对齐器
def get_face(fa, image):
    detector = dlib.get_frontal_face_detector()  # 获取人脸检测器
    gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)  # 将图像转换为灰度图
    thresh = gray.shape[0] // 4  # 设置阈值
    rects = detector(gray, 2)  # 检测人脸
    face_aligned = None  # 初始化返回的人脸图像
    for rect in rects:
        (x, y, w, h) = rect_to_bb(rect)  # 获取人脸的坐标
        if w > thresh:  # 如果人脸宽度大于阈值，则认为是有效人脸
            face_aligned = fa.align(image, gray, rect)  # 对齐人脸
            break  # 只处理第一张人脸
    return face_aligned


# 处理视频
def process_video(video_path, save_dir, fa):
    cap = cv2.VideoCapture(video_path)  # 打开视频文件
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))  # 获取总帧数
    if total_frames < 64:  # 如果视频帧数少于64，跳过该视频
        print(f"Warning: Video '{video_path}' has less than 64 frames. Skipping.")
        cap.release()  # 释放视频文件
        return

    start_frame = random.randint(0, total_frames - 64)  # 随机选择起始帧
    frames = []
    for i in range(start_frame, start_frame + 64):  # 提取连续的64帧
        cap.set(cv2.CAP_PROP_POS_FRAMES, i)  # 设置当前读取的帧数
        ret, frame = cap.read()  # 读取该帧
        if ret:
            frames.append(frame)  # 保存读取到的帧

    cap.release()  # 释放视频文件

    for i, frame in enumerate(tqdm(frames, desc=f"Processing frames from {os.path.basename(video_path)}")):  # 加入进度条
        face_aligned = get_face(fa, frame)  # 对齐每一帧中的人脸
        if face_aligned is not None:
            img_name = f"{i + 1:05d}.jpg"  # 给每一帧命名
            save_path = os.path.join(save_dir, img_name)  # 保存路径
            cv2.imwrite(save_path, face_aligned)  # 保存图像
        else:
            print(f"Face not found in frame {i + 1}")  # 如果没有检测到人脸


# 主函数：处理数据集中的所有视频
def align_dlib():
    predictor = dlib.shape_predictor(r"../weights/shape_predictor_68_face_landmarks.dat")  # 加载预测器
    fa = FaceAligner(predictor, desiredFaceWidth=crop_size)  # 初始化人脸对齐器

    # 遍历主目录（Training、Development、Testing）
    main_dirs = ['Testing']
    for main_dir in main_dirs:
        main_dir_path = os.path.join(dataset_path, main_dir)
        if not os.path.isdir(main_dir_path):
            print(f"Skipping non-directory: {main_dir_path}")
            continue

        # 遍历每个子目录（Northwind、Freeform 等）
        sub_dirs = os.listdir(main_dir_path)
        # for sub_dir in sub_dirs:
        #     sub_dir_path = os.path.join(main_dir_path, sub_dir)
        #     if not os.path.isdir(sub_dir_path):
        #         print(f"Skipping non-directory: {sub_dir_path}")
        #         continue

        # 遍历视频文件夹中的每个视频文件
        video_files = os.listdir(main_dir_path)
        for video_file in video_files:
            video_path = os.path.join(main_dir_path, video_file)

            if not os.path.isfile(video_path):
                continue

            # 获取视频名称（去掉文件扩展名）
            video_name = os.path.splitext(video_file)[0]

            # 构建保存路径： datasets/avec14/Training/Northwind/236_1_Northwind_video
            save_path = os.path.join(output_path, main_dir, video_name)
            os.makedirs(save_path, exist_ok=True)  # 创建保存文件夹

            print(f"Processing video: {video_path}")
            process_video(video_path, save_path, fa)  # 处理该视频


if __name__ == "__main__":
    align_dlib()  # 调用主函数进行处理

debug

eyesCenter在cv2.getRotationMatrix2D中一些版本要求是传入float型，直接传整型可能报错：

Traceback (most recent call last):
File “D:\python_project\face-parsing\utils\face_dect.py”, line 99, in
align_dlib() # 调用主函数进行处理
File “D:\python_project\face-parsing\utils\face_dect.py”, line 95, in align_dlib
process_video(video_path, save_path, fa) # 处理该视频
File “D:\python_project\face-parsing\utils\face_dect.py”, line 49, in process_video
face_aligned = get_face(fa, frame) # 对齐每一帧中的人脸
File “D:\python_project\face-parsing\utils\face_dect.py”, line 24, in get_face
face_aligned = fa.align(image, gray, rect) # 对齐人脸
File “C:\Users\Fine\anaconda3\envs\torch2\lib\site-packages\imutils\face_utils\facealigner.py”, line 68, in align
M = cv2.getRotationMatrix2D(eyesCenter, float(angle), float(scale))
TypeError: Can’t parse ‘center’. Sequence item with index 0 has a wrong type

将人脸对齐脚本中的eyesCenter类型转换为float即可：

		# eyesCenter = ((leftEyeCenter[0] + rightEyeCenter[0]) // 2,
		# 	(leftEyeCenter[1] + rightEyeCenter[1]) // 2)
		eyesCenter = ((leftEyeCenter[0] + rightEyeCenter[0]) / 2.0,
					  (leftEyeCenter[1] + rightEyeCenter[1]) / 2.0)

		# grab the rotation matrix for rotating and scaling the face
		M = cv2.getRotationMatrix2D(eyesCenter, float(angle), float(scale))

参考：
LinlyZhai-对AVEC2014视频进行Dlib或MTCNN人脸裁剪