Research On Face Detection Algorithm Generated By Local GAN Based On Convolutional Neural Networ

In recent years,with the rapid development of deep learning technology and the increasing popularity of image editing tools,People can not only forge the whole face by face editing software or Generative Adversarial Network(GAN),but also use GAN-based deep inpainting models to generate brand-new semantic content in the local area of face,i.e,locally GANgenerated face.The emergence of a large number of fake faces has brought great challenges to reputation protection,judicial system and even social stability.Therefore,it is urgent to propose effective methods to detect these fake faces.Aiming at locally GAN-generated faces,the existing detection algorithms have the following two problems:(1)The generated faces are often accompanied by a variety of postprocessing operations in the process of generation and propagation,but the existing algorithms are not robust enough.(2)Although the existing algorithms have good detection performance on in-dataset test,but their generalization still needs to be improved.Therefore,this thesis will carry out the following two aspects of research work on the above problems:(1)Aiming at the problem of insufficient robustness of locally GAN-generated face detection algorithm,this thesis proposes a robust locally GAN-generated face detection algorithm based on multi-level dual-stream features fused by attention.The algorithm takes into account the more stable characteristics of the abnormal traces of GAN-generated images in RGB color space and YCb Cr color space,a dual-stream network based on Xception network is designed to extract robust features from these two color spaces respectively.Then,the attention feature fusion module is used to fuse the dual-stream features on different network layers to further improve the robustness.In order to enhance the performance of the network in extracting and discriminating the features of the locally generated region,the decision is made by fusing multi-level features.Experimental results show that the proposed algorithm is generally better than existing algorithms in terms of the detection accuracy and robustness.(2)Aiming at the problem of insufficient generalization of locally GAN-generated face detection algorithm,this thesis proposes an adaptive frequency-aware general locally GANgenerated face detection algorithm.Since the up-sampling operation in GAN will leave frequency artifacts in the generated image,the proposed algorithm designs an adaptive frequency-aware module consisting of a high-pass filter and a learnable filter to highlight the abnormal frequency artifacts.Then the Xception network is used as the detection network,and the 8 residual blocks in Xception are deleted and the ECA attention mechanism is added to the remaining residual blocks to enhance its ability to capture artifacts.In addition,based on the Celeb A-HQ face dataset,this thesis uses four GAN-based deep inpainting models to construct diversified locally GAN-generated face dataset for generalization experiments.Experimental results show that the proposed algorithm is better than the existing algorithms in terms of the generalization,and it is robust to Gaussian noise.