Compressed Face Image Restoration Method Based On Deep Learning

Lossy compressed image restoration is a classical problem in the field of image processing.Over the years,many studies on compressed image restoration have yielded significant results.However,most compressed image restoration algorithms are for natural images,and the particularity of facial image structure is not specifically considered.Therefore,the restoration result of compressed facial images is not ideal.Aiming at this problem,this paper proposes a compressed face image restoration algorithm based on deep learning combining the prior information of facial components.The main research results are as follows:1.Facial structure based compressed face image restoration algorithm with deep residual networks.Due to the specific properties of the domain,the use of a universal compressed image restoration algorithm for compressed face image restoration often results in severe blur near the facial components.In order to obtain a clearer restored facial components,this paper proposes a deep residual network combining the prior information of facial structure.In the training phase,a pre-recovery network for initial restoration of the overall structure of the face image is first trained.However,the pre-recovery network without any prior information cannot handle the facial components with fine structure well.Therefore,five facial components,including eyes,eyebrows,nose,mouth and face profile,are extracted based on the pre-recovery network,respectively,for training the detail enhancement networks for restoring the facial components.At the restoration stage,combining with the facial mask generated by detection results at key points,the corresponding network parameters are used for accurate restoration of different facial components.The experimental results show that this method not only has a higher evaluation index than other restoration algorithms,but also has clearer texture details at the facial components.2.Compressed face image restoration algorithm based on generative adversarial networks.The model optimized using MSE criteria usually has a high signal-to-noise ratio,but the resulting images tend to be smooth and have poor visual perception.Perceptual loss can restore the semantic structure of the image and generate an image with excellent perceptual quality.Therefore,this paper proposes a method of optimizing generative adversarial networks using a perceptual loss function for compressed face image restoration.If the compressed image is used directly to generate adversarial training,the generator may misidentify the original noise in the low-quality image as the details of the image itself and zoom in.Therefore,this paper first trains a denoising model that combines facial components to remove excess noise in the input image.Then it uses a generative adversarial network to generate textures and details at the facial components.The experimental results show that the denoising model achieves the best PSNR and SSIM,and the generation of the adversarial network model generates a richer face image with the best perceptual quality.