代码解析—part1 数据准备—CVPR2023—Implicit Identity Leakage: The Stumbling Block to Improving Deepfake

DeepLearning01

已于 2024-05-23 15:17:00 修改

阅读量371

点赞数 4

分类专栏： AIGC生成论文文章标签：学习人工智能视觉检测 AIGC

于 2024-05-13 00:05:13 首次发布

本文链接：https://blog.csdn.net/JustWantToLearn/article/details/138773005

版权

AIGC生成论文专栏收录该内容

8 篇文章 0 订阅

订阅专栏

论文讲解请看：https://blog.csdn.net/JustWantToLearn/article/details/138758033
代码链接：https://github.com/megvii-research/CADDM
在这里，我们简要描述算法流程，着重分析模型搭建细节，以及为什么要这样搭建。
part 1：数据集准备本文
part 2: 数据集加载，包含 Multi-scale Facial Swap(MFS) 模块：https://blog.csdn.net/JustWantToLearn/article/details/139092687
part 3：训练过程，ADM模块 https://blog.csdn.net/JustWantToLearn/article/details/139116455

环境准备

linux
Python 3 >= 3.6
Pytorch >= 1.6.0
OpenCV >= 4.4.0
Scipy >= 1.4.1
NumPy >= 1.19.5

在安装dlib库的时候卡了很久，先安装cmake，再去安装dlib。如果不成功，直接源码安装

git clone https://github.com/davisking/dlib.git
cd dlib
mkdir build
cd build
cmake ..
cmake --build . --config Release
cmake --build . --config Release --target install
cd ..
python setup.py install

数据准备提取ff++数据集的ldm

1、下载FF++数据集

根据链接https://github.com/ondyari/FaceForensics/tree/master提示下载
放入特定文件夹中

.
└── data
    └── FaceForensics++
        ├── original_sequences
        │   └── youtube
        │       └── raw
        │           └── videos
        │               └── *.mp4
        ├── manipulated_sequences
        │   ├── Deepfakes
        │       └── raw
        │           └── videos
        │               └── *.mp4
        │   ├── Face2Face
        │		...
        │   ├── FaceSwap
        │		...
        │   ├── NeuralTextures
        │		...
        │   ├── FaceShifter
        │		...

2、提取ldm

 python lib/extract_frames_ldm_ff++.py

2.1 定义路径，函数

#extract_frames_ldm_ff++.py
#!/usr/bin/env python3
from glob import glob
import os
import cv2
from tqdm import tqdm
import numpy as np
import dlib
import json
import argparse
from imutils import face_utils

#定义路径
VIDEO_PATH = "./CADDM/data/FaceForensics++"
SAVE_IMGS_PATH = "./test_images"
#下载好的形状预测模型
PREDICTOR_PATH = "./CADDM/lib/shape_predictor_81_face_landmarks.dat"
DATASETS = {'Original', 'FaceSwap', 'FaceShifter', 'Face2Face', 'Deepfakes', 'NeuralTextures'}
COMPRESSION = {'raw'}
NUM_FRAMES = 1
IMG_META_DICT = dict()

def parse_labels(video_path):
    label = None
    if "original" in video_path:
        label = 0
    else:
        label = 1
    return label

def parse_source_save_path(save_path):
    source_save_path = None
    if "original" in save_path:
        source_save_path = save_path
    else:
        img_meta = save_path.split('/')
        source_target_index = img_meta[-1]
        source_index = source_target_index.split('_')[0]
        manipulation_name = img_meta[-4]
        original_name = "youtube"
        source_save_path = save_path.replace(
            "manipulated_sequences", "original_sequences"
        ).replace(
            manipulation_name, original_name
        ).replace(
            source_target_index, source_index
        )
    return source_save_path

2.2 main函数主函数


def main():
	# 初始化dlib的面部检测器（HOG模型）
    face_detector = dlib.get_frontal_face_detector()
    # 加载形状预测模型
    face_predictor = dlib.shape_predictor(PREDICTOR_PATH)
    #遍历文件{'Original', 'FaceSwap', 'FaceShifter', 'Face2Face', 'Deepfakes', 'NeuralTextures'}
    for dataset in DATASETS:
        for comp in COMPRESSION:
            # 加载文件夹下所有mp4文件
            movies_path_list = parse_video_path(dataset, comp)
            n_sample = len(movies_path_list)
            for i in tqdm(range(n_sample)):
                #使用 replace 方法生成保存视频帧的路径
                save_path_per_video = movies_path_list[i].replace(
                    VIDEO_PATH, SAVE_IMGS_PATH
                ).replace('.mp4', '').replace("/videos", "/frames")
                preprocess_video(
                    movies_path_list[i], save_path_per_video,
                    face_detector, face_predictor
                )
    with open(f"{SAVE_IMGS_PATH}/ldm.json", 'w') as f:
        json.dump(IMG_META_DICT, f)


if __name__ == '__main__':
    main()
# vim: ts=4 sw=4 sts=4 expandtab

2.3 parse_video_path函数读取所有video

def parse_video_path(dataset, compression):
    # this path setting follows FF++ dataset
    if dataset == 'Original':
        dataset_path = f'{VIDEO_PATH}/original_sequences/youtube/{compression}/videos/'
    elif dataset in ['FaceShifter', 'Face2Face', 'Deepfakes', 'FaceSwap', 'NeuralTextures']:
        dataset_path = f'{VIDEO_PATH}/manipulated_sequences/{dataset}/{compression}/videos/'
    else:
        raise NotImplementedError
    # get all videos under the specific manipulated/original sequences
    #使用 glob 模块查找所有以 .mp4 结尾的视频文件
    movies_path_list = sorted(glob(dataset_path+'*.mp4'))
    print("{} : videos are exist in {}".format(len(movies_path_list), dataset))
    return movies_path_list

2.4 preprocess_video函数，检测面部并保存地标

def preprocess_video(video_path, save_path, face_detector, face_predictor):
    # save the video meta info here
    video_dict = dict()
    # get the labels
    #真实视频标签为0，假视频标签1
    label = parse_labels(video_path)
    # 获取原始视频id，如果造假视频id   110_043,则source id 110，原始视频不变
    #'./manipulated_sequences/FaceSwap/raw/frames/110_043' --》'./original_sequences/youtube/raw/frames/110'
    source_save_path = parse_source_save_path(save_path)
    # prepare the save path
    os.makedirs(save_path, exist_ok=True)
    # read the video and prepare the sampled index
    #使用OpenCV的VideoCapture类打开视频文件
    cap_video = cv2.VideoCapture(video_path)
    #获取视频的总帧数
    frame_count_video = int(cap_video.get(cv2.CAP_PROP_FRAME_COUNT))
    #使用numpy.linspace生成一个等间隔的整数数组，表示要提取的帧的索引，NUM_FRAMES为1
    #这里只保存了每个视频第一帧图片
    frame_idxs = np.linspace(0, frame_count_video - 1, NUM_FRAMES, endpoint=True, dtype=int)
    # process each frame
    # 循环遍历视频的每一帧
    for cnt_frame in range(frame_count_video):
        # cap_video.read() 读取当前帧并返回读取状态ret和帧数据frame
        ret, frame = cap_video.read()
        height, width = frame.shape[:-1]
        if not ret:
            tqdm.write('Frame read {} Error! : {}'.format(cnt_frame, os.path.basename(video_path)))
            continue
        # 检查当前帧索引是否在要提取的帧索引数组中。如果不是，跳过此帧
        if cnt_frame not in frame_idxs:
            continue
        # 将当前帧从BGR颜色空间转换为RGB颜色空间，以便与dlib的面部检测器兼容
        frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        # 使用dlib的面部检测器检测当前帧中的面部，1 表示检测器将图像进行上采样一次，以检测更小的面部
        faces = face_detector(frame, 1)
        if len(faces) == 0:
            tqdm.write('No faces in {}:{}'.format(cnt_frame, os.path.basename(video_path)))
            continue
        # 一个用于保存面部特征点，另一个用于保存检测到的面部大小
        landmarks = list()  # save the landmark
        size_list = list()  # save the size of the detected face
        for face_idx in range(len(faces)):
            landmark = face_predictor(frame, faces[face_idx])
            landmark = face_utils.shape_to_np(landmark)
            x0, y0 = landmark[:, 0].min(), landmark[:, 1].min()
            x1, y1 = landmark[:, 0].max(), landmark[:, 1].max()
            face_s = (x1 - x0) * (y1 - y0)
            size_list.append(face_s)
            landmarks.append(landmark)
        # save the landmark with the biggest face
        #若一张图片上出现多张人脸，选择人脸特征最大的脸
        # 将所有面部特征点数组连接成一个NumPy数组，并根据面部大小排序，选择最大的面部特征点。
        landmarks = np.concatenate(landmarks).reshape((len(size_list),)+landmark.shape)
        landmarks = landmarks[np.argsort(np.array(size_list))[::-1]][0]
        # save the meta info of the video
        # 将最大的面部特征点、帧路径和标签保存到video_dict字典中
        video_dict['landmark'] = landmarks.tolist()
        video_dict['source_path'] = f"{source_save_path}/frame_{cnt_frame}"
        video_dict['label'] = label
        IMG_META_DICT[f"{save_path}/frame_{cnt_frame}"] = video_dict
        # save one frame
        # 将帧从RGB颜色空间转换回BGR颜色空间，构建保存帧图像的路径
        frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
        image_path = f"{save_path}/frame_{cnt_frame}.png"
        cv2.imwrite(image_path, frame)
    cap_video.release()
    return