Pixel-wise Contrastive Learning For Image Super-Resolution

Image super-resolution reconstruction refers to the use of low resolution single images or a group of images to obtain high-resolution images of corresponding images,thereby improving image quality.Deep convolutional neural networks have dominated the current research on single image super-resolution(SISR)technology and made significant progress.However,SISR is still an open issue,and reconstructed super-resolution(SR)images often encounter issues such as blurring,loss of texture details,and distortion.In this paper,a new pixel-wise contrastive loss has been proposed.In a local area,keeping the SR pixels as close as possible to the original HR pixels and away from other pixels in the local area,which can significantly improve the fidelity and visual quality of the SR image.We also proposes an progressive residual feature fusion SR network architecture that combines with the comparative loss.Our main contributions include:(1)A general pixel-wise contrastive loss function based on contrastive learning is proposed,which can improve the fidelity and visual quality of SR images.In image space,contrastive loss brings the SR pixels as close as possible to the original HR pixels and away from other pixels in the local area.Also the loss function can be used in collaboration with other SR network architectures.(2)A lightweight multi-scale residual channel attention block is proposed to better fuse multi-scale features.(3)A spatial attention fusion block is proposed,which can better fuse spatial features.(4)A progressive residual feature fusion SR network architecture is proposed,combining a multi-scale residual channel attention block,a spatial attention fusion block,and a pixel-by-pixel contrast loss.The experimental results show that competitive SR performance is achieved,which is significantly better than other representative methods.