English

应用科学学报 ›› 2019, Vol. 37 ›› Issue (5): 618-630.doi: 10.3969/j.issn.0255-8297.2019.05.004

• 多媒体信息安全 • 上一篇    下一篇

基于卷积神经网络的面部图像修饰检测

王灿军1,2, 廖鑫1,3, 陈嘉欣1,2, 秦拯1,2, 刘绪崇3   

  1. 1. 湖南大学 信息科学与工程学院, 长沙 410012;
    2. 湖南大学 大数据研究与应用湖南省重点实验室, 长沙 410082;
    3. 湖南警察学院 网络犯罪侦查湖南省重点实验室, 长沙 410138
  • 收稿日期:2019-07-27 修回日期:2019-08-01 出版日期:2019-09-30 发布日期:2019-10-18
  • 通信作者: 廖鑫,副教授,博导,研究方向:多媒体安全与取证,E-mail:xinliao@hun.edu.cn E-mail:xinliao@hun.edu.cn
  • 基金资助:
    国家自然科学基金（No.61972142，No.61772191）；网络犯罪侦查湖南省普通高校重点实验室开放课题（No.2017WLFZZC001）资助

Research on Facial Modification Detection Algorithm Based on Convolutional Neural Network

WANG Canjun1,2, LIAO Xin1,3, CHEN Jiaxin1,2, QIN Zheng1,2, LIU Xuchong3   

  1. 1. College of Computer Science and Electronic Engineering, Hunan University, Changsha 410012, China;
    2. Hunan Key Laboratory of Big Data Research and Application, Hunan University, Changsha 410082, China;
    3. Hunan Key Laboratory of Cybercrime Reconnaissance, Hunan Police College, Changsha 410138, China
  • Received:2019-07-27 Revised:2019-08-01 Online:2019-09-30 Published:2019-10-18

PDF

113

摘要： 为避免人为因素对人脸面部图像皮肤纹理特征提取产生的影响，用卷积神经网络算法对人脸图像修饰进行检测.传统的图像分类方法需要进行复杂的人工特征提取，而卷积神经网络可以自动学习并直接从图像中获取特征，解决了传统模式识别方法特征提取难的问题，具有更高的识别率和更广泛的实用性.在传统卷积神经网络模型中，调整卷积核大小、减少参数、改变卷积层滤波器数量、调整卷积层和池化层的交替方式、使用dropout来提高模型泛化能力以形成适用于人脸修饰检测的新的网络模型.实验结果表明，在引入的数据集上，新的网络模型对人脸图像的修饰检测有较强的鲁棒性，达到了较高的识别率.

关键词: 卷积神经网络, 深度学习, 人脸图像修饰, 修饰检测, 纹理特征

Abstract: In order to avoid the influence of human factors on skin texture feature extraction of facial images, this paper proposes to detect facial image modification by using convolutional neural network (CNN) algorithm. To the best of our knowledge, this is the first report to use CNN in the detection of human face tampering. Compared with the traditional image classification methods which need complex artificial feature extraction, CNN can learn automatically, acquire features directly from the image, and reduce the difficulty of extracting features in traditional pattern recognition methods, accordingly, gaining a higher recognition rate and wider practicality at the same time. On the basis of the traditional convolutional neural network model, the proposed method builds a new network model for human face tampering detection by adjusting the size of the convolution kernel, reducing the parameters, changing the number of convolutional layer filters, adjusting the alternate mode of the convolutional layer and the pooling layer, and using dropout to improve the generalization ability of the model. Experimental results show that the new network model performs with a high recognition rate and strong robustness in the tamper detection of facial images.

Key words: convolutional neural network (CNN), deep learning, facial image retouching, retouching detection, texture feature

中图分类号: