Research On Deep Learning-based Face Image Super-Resolution Algorithms

As the most basic and most effective information of personal identity,face images are self-evident in the links of public security organ clue extraction,action route tracking and suspect identity information confirmation.In the process of face image acquisition,in the face of special circumstances,such as the performance of the monitoring equipment itself or the influence of complicated weather,the details of the acquired face image are blurred,and the resolution is low.Because of the small difficulty,the research on super-resolution reconstruction technology of face images has attracted the attention of many experts and scholars.Because of the small difficulty,the research on super-resolution reconstruction technology of face images has attracted the attention of many experts and scholars.Although the existing face image super-resolution reconstruction technology can restore the low-resolution face image to a certain extent,it still does not achieve a very ideal effect.Therefore,this paper conducts in-depth research on the basis of the generative confrontation network model that is applied to face super-resolution reconstruction,and proposes two improved reconstruction algorithms for its shortcomings in face image reconstruction.The main contents of this article are as follows:1)First,it gives a detailed overview of the basic background of image super-resolution reconstruction technology.At the same time,it explains the theory of face image super-resolution reconstruction method and analyzes the existing reconstruction methods,and summarizes the advantages and disadvantages that still exist.The subjective and objective evaluation criteria of the reconstructed face image quality are introduced,and the subjective and objective evaluation methods that will be applied to the research of this article are determined.2)An improved face image super-resolution reconstruction algorithm based on SRGAN is proposed.Aiming at the drawbacks of the Generative Adversarial Network(GAN)model in the super-resolution reconstruction of face images,three improvements are proposed,including: the introduction of the self-attention mechanism to strengthen the neural network to capture the global feature dependence of the face image ability;Replace the original normalization layer to ensure network convergence speed and stability,while avoiding damage to the spatial representation of the original image;Introduce PatchGAN idea,strengthen and improve the ability of texture detail discrimination,improve the quality of reconstruction,compared with the original the SRGAN discriminator reduces the calculation parameters for the whole image discrimination and improves the resource utilization.Finally,a comparative test with other traditional algorithms is performed on the public face data set CelebA,and the subjective and objective evaluation criteria are integrated to verify the effectiveness of the method.3)LapGAN reconstruction algorithm based on improved SRGAN segmentation processing is proposed.Regarding the existing algorithms and improved SRGAN,when reconstructing large-scale face images,they contain a lot of information,resulting in a large amount of calculations and difficulty in training process when reconstructing large-scale and multi-feature information in a single time.The multi-level pyramid model structure is proposed,and a segmented face super-resolution reconstruction method is proposed based on the LapGAN model.The generator and discriminator network are designed,and the loss function is improved.Finally,a comparative test with other traditional algorithms is performed on the face data set composed of CelebA and LFW,and the subjective and objective evaluation criteria are integrated to effectively verify the segmented reconstruction ability and effect of the method.