Research On Super-resolution Feature Reconstruction Of Face Images Based On Deep Learnin

Super-resolution algorithms are image and video restoration algorithms that work by taking a low-resolution input image and outputting a high-resolution image while reconstructing high-frequency details as much as possible.Face image super-resolution is a subtask of image super-resolution.There are various types of prior information in face images,such as landmark prior information,heat maps,and parsing maps.However,existing methods still have shortcomings in the application of prior information.Therefore,this paper proposes a face image super-resolution convolutional neural network based on prior information to improve the reconstruction effect.On the other hand,in addition to the realism of the generated images,the quality of the visual perception of the images is also an important factor in evaluating the algorithm.This paper proposes a scheme for face image super-resolution using a diffusion model to address this aspect.The main work of this paper is as follows:To balance the pixel loss and perceptual of the reconstruction result,this work proposes a convolutional neural network for face image super-resolution based on prior information(Global Attention Guided Multi-scale Network,GAGMN).The network consists of three sub-modules: the upsampling module,the prior estimation module,and the image reconstruction module.The upsampling module increases the resolution and reconstructs the coarse super-resolution result.The prior prediction module predicts the prior information based on the output of the upsampling module,which is used for the next step of fusion reconstruction.The image reconstruction module fuses the output results of the first two modules to achieve the final super-resolution reconstruction result.To improve the output’s perceptual quality and consider the generated results’ reference objective evaluation metrics,this paper proposes a Diffusion Module Based Face Image Super-resolution Network(DMBN).This paper applies the diffusion model network to the task of super-resolution of face images and designs the most critical part of the model,the noise prediction network.For the structure of the noise prediction network,a dual-stream network composed of Transformer and U-Net branches is designed for noise prediction.The improved noise prediction network can effectively predict noise based on the denoising process.Based on the predicted noise results,the diffusion model network can obtain the final reconstruction result through step-by-step denoising.This paper validates the proposed methods on public datasets and compares them with other algorithms through quantitative and qualitative experiments,demonstrating the effectiveness of the proposed approaches.The results show that the two methods can be tailored to different application scenarios.