Image Super-resolution With Improved Perceptual Quality

Image super-resolution(SR)is one of the classical research contents in computer vision,which is the task of recovering a high-resolution(HR)image from low-resolution(LR)images.The goal of SR is to produce results with high reconstruction accuracy and high visual quality.In real scenarios,it is necessary to recover realistic image details with a largescale factor.In recent years,the development of the deep convolutional neural network(CNN)promotes the improvement of SR performance.This paper focuses on improving SR images perceptual quality.After fully studying the existing deep learning SR methods,we proposed a generative adversarial network(GAN)with multi-feature discriminators for SR.It effectively improves SR images perceptual quality.In order to further improve the perceptual quality of feature-rich regions in the image,the weighted content loss function is proposed.Numerous experimental results show that the GAN with multi-feature discriminator optimized by the weighted content loss function can effectively improve the visual quality of feature-rich regions and produce visually pleasing HR results.The work is mainly reflected in the following three aspects:(1)Common SR methods,mainly including interpolation-based methods,shallow learningbased methods,and deep learning-based methods are studied,and perceptual quality-driven deep learning-based SR methods are focused.Several representative methods are implemented and analyzed.The SR results with high perceptual quality have important demand value in practice.(2)In this paper,the generative adversarial network framework with multi-feature discriminators(MFDGAN)for SR is proposed.MFDGAN includes generator,image discriminator,morphological component discriminator,and color discriminator.The adversarial learning between the generator and multi-feature discriminator forces the edge,texture and color information of the SR image to be consistent with that of the HR image.Experimental results show that our method effectively enhances the edge and texture details in the reconstructed image and avoid color distortion.(3)A new weighted content loss function is proposed to optimize the generative adversarial network framework with multi-feature discriminators(MFDGANW).Considering that vision is sensitive to the feature-rich regions,the weighted content loss function makes the optimized GAN with multi-feature discriminator recover feature-rich regions in the SR image.A large amount of experimental results shows that MFDGANW can greatly improve the perceptual quality of SR images reconstructed from different kinds of LR images with low computational complexity.The SR results are pleasant in the edge,texture,color,and feature-rich regions.