Research And Application Of Optimization Method For Super-Resolution Reconstruction Of Face Images

The extremely low-resolution face image can only provide people with less information,which brings challenges to the subsequent face detection and recognition technology.Face image super-resolution reconstruction technology can obtain a high-resolution image by processing one or more low-resolution images,thereby providing a good basis for subsequent processing.Currently,a more mature 4x face image reconstruction technology has been developed.However,in real application scenarios,a larger factor reconstruction is required for very lowresolution face images.However,when the reconstruction factor is 8,the effectiveness of most reconstruction techniques will drop quickly.Therefore,the main goal and work of the thesis is to use deep learning technology to improve the effect of 8x reconstruction of extremely lowresolution face images.The specific research content of the thesis is as follows:(1)Research and design of face super-resolution reconstruction method based on residual blocks.The thesis is based on classical reconstruction model SRResNet,which is based on residual blocks,and improves its residual blocks from the perspective of the utilization of two features,the model FRSENet based on residual SE module and the model FRDBNet based on residual dense block are constructed respectively.One feature utilization angle is feature recalibration—values are assigned to each characteristic channel of face in order to encourage residual blocks to learn features that are beneficial to reconstruction effect,specifically by building a residual SE module;Another angle is feature reuse—densely connected convolution layers share features from front to back to enrich the features in learning process and improve learning effect.Experiments show that the two proposed reconstruction models can improve the reconstruction effect.Among them,FRSENet's value of PSNR and SSIM are improved by 0.19 db and 1% respectively compared with SRResNet;FRDBNet's value of PSNR and SSIM are improved by 0.29 db and 1.02% respectively compared with SRResNet,showing that the reconstruction effect of model FRDBNet is better.(2)Research and design of face super-resolution reconstruction method based on progressive upsampling network.Based on the research results of model FRDBNet,the thesis adopts a structure of progressive upsampling network—The commonly used method of upsampling at the end of the network is changed to multi-stage upsampling,and a large multiple reconstruction is achieved by up-sampling stages with lower multiple reconstruction task.At the same time,using this structure to compute loss function between the label images and the reconstructed images at different stages to replace the commonly used loss function,achieving multi-stage effective supervision of network parameter optimization.Finally,in order to further restore the highfrequency information of the face image,the thesis introduces a generative adversarial training mechanism and constructs a face reconstruction model EFRDBNet-GAN based on generative adversarial network.The experimental results show that values of PSNR and SSIM of EFRDBNet are improved by 0.45 db and 1.05% respectively compared to SRResNet,and EFRDBNet-GAN provides the best visual experience.(3)Design of prototype system for face super-resolution reconstruction.The thesis uses Python language to realize the face super-resolution reconstruction prototype system.The system mainly provides a test function for the reconstruction method proposed in the thesis,and at the same time realizes the management function of the reconstructed image,providing a good experience for users.