基于对比学习的深度残差网络图像超分辨率方法

Depth Residual Image Super-Resolution Based on Contrast Learning

传统的基于对比学习的图像超分辨率方法,一般将原始图像作为正样本,将退化图像或其他类图像作为负样本,存在对纹理细节恢复差的问题。本文提出基于对比学习的深度残差网络图像超分辨率(depth residual image super-resolution based on contrast learning,CEDSR)方法,针对残差超分辨率模型,采用对高分辨率图像锐化后的图像作为正样本,对高分辨率图像轻微模糊的图像作为负样本,利用正负样本下的对比损失提升对纹理细节的恢复增强。增强锐化后的正样本图像携带更丰富的纹理信息,基于不同函数生成的模糊负样本图像刻画了纹理模糊特征,正负样本构建的对比损失有利于图像超分辨率图像对纹理细节的恢复。本文模型在4个标准数据集DIV2K、Set14、BSDS100和Urban100上与经典算法进行实验对比,定性和定量实验结果均表明本文模型可以获得效果更好的超分辨率图像。

The traditional image super-resolution method based on contrast learning generally takes the original image as the positive sample and the degraded image or other types of images as the negative sample,which has the problem of poor texture detail restoration.In this article,we propose a depth residual image super-resolution based on contrast learning(CEDSR)method.In our proposed method,for the residual super-resolution model,the sharpened images of high-resolution images are used as positive samples,and the slightly blurred images of high-resolution images are used as negative samples.The contrast loss lifting under positive and negative samples is used to restore and enhance the texture details.The positive sample images after enhancement and sharpening carry more abundant texture information,and the fuzzy negative sample images generated based on different functions depict the texture fuzzy features.The contrast loss of positive and negative samples is conducive to the restoration of texture details in super-resolution images.Our proposed model was compared with the classical algorithms on DIV2K,Set14,BSDS100 and Urban100 standard data sets.The qualitative and quantitative experimental results show that the model has better super-resolution image.