基于像素对比学习的图像超分辨率算法

目前, 深度卷积神经网络(Convolutional neural network, CNN)已主导了单图像超分辨率(Single image super-resolution, SISR)技术的研究, 并取得了很大进展. 但是, SISR仍是一个开放性问题, 重建的超分辨率(Super-resolution, SR)图像往往会出现模糊、纹理细节丢失和失真等问题. 提出一个新的逐像素对比损失, 在一个局部区域中, 使SR图像的像素尽可能靠近对应的原高分辨率(High-resolution, HR)图像的像素, 并远离局部区域中的其他像素, 可改进SR图像的保真度和视觉质量. 提出一个组合对比损失的渐进残差特征融合网络(Progressive residual feature fusion network, PRFFN). 主要贡献有: 1)提出一个通用的基于对比学习的逐像素损失函数, 能够改进SR图像的保真度和视觉质量; 2)提出一个轻量的多尺度残差通道注意力块(Multi-scale residual channel attention block, MRCAB), 可以更好地提取和利用多尺度特征信息; 3)提出一个空间注意力融合块(Spatial attention fuse block, SAFB), 可以更好地利用邻近空间特征的相关性. 实验结果表明, PRFFN显著优于其他代表性方法.

Pixel-wise Contrastive Learning for Single Image Super-resolution

Deep convolutional neural network (CNN) has achieved great success in single image super-resolution (SISR). However, SISR is still an open issue, and reconstructed super-resolution (SR) images often suffer from blurring, loss of texture details and distortion. In this paper, a new pixel-wise contrastive loss is proposed to improve the fidelity and visual quality of SR images by making the pixels of SR images as close as possible to the corresponding pixels of the original high-resolution (HR) images and away from the other pixels in the local region. We also propose a progressive residual feature fusion network (PRFFN) with combined contrastive loss, and the main contributions include: 1) A general pixel-wise loss function based on contrastive learning is proposed, which can improve the fidelity and visual quality of SR images; 2) A lightweight multi-scale residual channel attention block (MRCAB) is proposed, which can better extract and utilize multi-scale feature information; 3) A spatial attention fusion block (SAFB) is proposed, which can better utilize the correlation of neighboring spatial features. The experimental results demonstrate that PRFFN significantly outperforms other representative methods.