Research On Face Image Quality Evaluation And Authenticity Identification Method Based On Multi-dimensional Features

With the deepening of the research on Generative Adversarial Network(GAN,Generative Adversarial Network),the technology of GAN to generate images has also made great progress.GAN technology can now easily generate fake face images that are fake enough,but for the human naked eye,it is extremely challenging to distinguish these GAN-generated face images.The forged face images generated by these GANs are likely to be used by some criminals to carry out some behaviors that endanger the society and bring great security risks to the society.With the advancement of GAN generation technology,the research on GAN-generated face image identification has also received more and more attention.It is imminent to find an effective and feasible GAN-generated face image identification technology.In the actual environment,the input image is not necessarily reliable,some lowquality face images are severely distorted,and the information contained in them is not enough to judge whether it is a fake image,or even some are not meaningful face images at all.The quality of the product has become one of the key factors affecting the identification of true and false.Therefore,it is very important to evaluate the quality of face images and determine whether they are suitable for GAN-generated face images.This paper conducts research on two aspects of face image identification and face quality evaluation generated by GAN.The main work is as follows:(1)Since color information and gradient domain texture information are of great significance to image quality evaluation,this paper proposes a face image quality evaluation model based on multi-dimensional image features based on the Transformer architecture.Different from the general convolutional neural network,the convolution operation in the convolutional neural network has a relatively limited receptive field of view,while the self-attention model can focus on the entire input sequence,so it can effectively capture image features of different granularities.Therefore,this paper uses a hash-based two-dimensional spatial embedding that maps patch locations to a fixed grid so that the Transformer model can leverage information across multiple dimensions.This paper compares it with other quality evaluation models on the GFIQA-20 k data set,and the experimental results prove the effectiveness and accuracy of the algorithm prediction proposed in this paper.(2)Aiming at the color information difference between real face images and GAN-generated face images in multiple color spaces,the face images are analyzed in multiple color spaces,and the gradient information and direction texture information are considered comprehensively.This paper proposes A GAN-generated face image identification method based on multi-dimensional features is proposed.In this paper,experiments are carried out on images under various perturbation attacks,and on various GAN datasets,and compared with other schemes.Experimental results show that the proposed method has good performance and is robust to various perturbation attacks.