Research On Super-resolution Reconstruction Of Human Face Images Based On Deep Learning

Face images are the most effective information to mark a person as an individual and can help the police to quickly obtain valuable information and play an important role in the investigation process.In real life,the acquired face images are affected by environmental interference,imaging equipment and other factors,and the acquired face images have problems such as blurred details and low resolution.Based on a generative adversarial network that performs well in face super-resolution reconstruction,this paper delves into its strengths and weaknesses in face image reconstruction and proposes two improved algorithms to address the shortcomings in order to achieve better reconstruction results.The main content of this article is as follows.1.Firstly,the theoretical background of the image super-resolution reconstruction method is outlined,and the existing face image super-resolution reconstruction methods are analyzed and the disadvantages that still exist are summarized.Next,we describe how to evaluate the quality of the reconstructed face images and identify the subjective and objective evaluation methods we used.Finally,the dataset used for the experiments in this paper is described.2.A super-resolution reconstruction algorithm for face images based on generative adversarial networks is proposed.In order to solve the problems of generative adversarial networks in processing face images,we propose some improvements,mainly including.Targeted extraction of high frequency information from face images by adding pixel and channel attention to the residual unit.Removal of the normalization layer of the residual unit to avoid corruption of the original image.Add regular term loss function to effectively remove noise from face images.The discriminative network constructed with Patch GAN idea can not only effectively enhance the recognition of texture details,but also significantly improve the quality of the reconstruction,which can greatly reduce the computational parameters compared with the traditional SRGAN discriminator,thus increasing the utilization of resources.Experimental comparisons with other algorithms on the publicly available datasets Celeb A,VGGFace2 and real face images demonstrate the effectiveness of the method by combining subjective and objective evaluation methods.3.Proposed a super-resolution reconstruction algorithm for face images based on pyramidal step-by-step processing.Because large scale face images contain rich detail information,existing algorithms generate a large amount of operations when reconstructing them,which makes the training complicated and difficult.Therefore,this chapter adopts the Lap GAN idea to establish a multi-level pyramid model structure,splitting the conventional 8-fold upsampling into a 3-stage 2-fold upsampling process,while replacing progressive upsampling with iterative upsampling layers,making feature extraction more adequate through iterative error correction,and improving the loss function to achieve efficient processing of face image data.After comparing experiments with other algorithms on the Celeb A dataset,the VGGFace2 dataset and real face images,combined with subjective and objective evaluation criteria,it can be proved that the method in this chapter outperforms other algorithms.