Super-resolution Reconstruction Of Mosaic Face Images Based On Deep Learning

In recent years,with the popularization of shooting equipment and image editing software,anyone can perform various operations on images,especially editing important face images,such as mosaic processing.Although this can bring certain benefits to the protection of fragile and private face information,it also brings certain difficulties to police forensics.Therefore,it is an urgent task to study to restore the face image after mosaic processing as much as possible.Among the various methods for mosaic processing of the face image content,downsampling and upsampling of the face image is the most common operation method.It refers to downsampling certain areas in the face image by a certain multiple,and then using the traditional upsampling method to restore the block to the original size,replacing this part of the image,so as to achieve the mosaic effect.Therefore,the research on the restoration algorithm of mosaic face images not only faces huge technical challenges,but also has strong practical application value.In this paper,in view of the above problems of face image mosaic technology and the advantages of deep learning technology in image processing,two different models are proposed to restore mosaic face images.Each of these two methods has its own advantages and insufficient.The main work and innovations of the paper are summarized as follows:First,in order to recover the original image from the mosaic face image,we propose an effective model,which we call DemosaicCNN.It combines the existing image super-resolution reconstruction models SRRes Net and RDN.We combine the advantages of the SRRes Net model and the RDN model to simplify the network structure by redesigning the Residual Dense Blocks,which is called Skip Residual Dense Blocks(SRDB).It can reduce the network parameters,improve the training speed,and can also combine the global features of the image with the local features of the feature map.The experimental results show that our proposed DemosaicCNN has achieved good results in super-resolution reconstruction of mosaic face images,especially the reconstructed images have higher PSNR values.However,the pictures generated by the DemosaicCNN network have problems of excessive smoothness and lack of texture details.In order to solve these problems,learn from the advantages of SRGAN and ESRGAN.Introducing GAN into our network,the Demosaic GAN model is proposed.The model also migrates Xception.Compared with VGG19,Xception reduces the parameters of the Demosaic GAN model.In addition,both the discriminator and the loss function are cleverly designed.The perceptual loss function is optimized by adding style loss,which makes the recovered image closer to the original image in high-dimensional features,and outputs a more realistic face image.The experimental results show that the restored mosaic face image is closer to the real face image,which solves the problem of excessive smoothness of the reconstructed image to a certain extent.