Abstract: In most convolutional neural networks-based super-resolution (SR) methods, the degradation assumptions are fixed and known (e.g., bicubic degradation). When applying them to real-world degraded images, the mismatch between the actual degradation and the assumptions can significantly degrade the performance of SR methods. To solve the problem of SR method performance degradation when degradation mismatch occurs, a new blind SR method is proposed, which can unsupervisedly learn the degradation information of degraded images and use the learned degradation information to efficiently modulate the SR network. Firstly, an unsupervised implicit degradation representation learning method is designed based on contrastive learning, which can extract accurate degradation information without relying on sample queues. It is also proposed to use different images that have gone through the same degradation as positive sample pairs, which overcomes the effect of image content similarity on the learned degradation, and effectively enhances the expressive power of the implicit degradation estimator. Subsequently, a degradation-guided modulation super-resolution network containing a global modulation block and a channel fusion modulation block is designed. It is combined with the degradation information and can flexibly deal with various kinds of degradations to realize the final SR recovery. The experiments are compared with various representative methods on four standard datasets, and the results show that the PSNR on Urban100 dataset is improved by 0.85 dB, 0.87 dB and 0.32 dB respectively over the baseline model DASR (degradation aware for blind super-resolution) for three different magnifications of ×2, ×3 and ×4. Both the quality of the final extracted degradation characterization information and reconstructed SR images are greatly improved over the DASR.

Key words: super-resolution, contrastive learning, degradation feature learning, degradation feature modulation

摘要： 大多数基于卷积神经网络的超分辨率（SR）方法，退化假设是固定且已知的（如双三次退化）。将它们应用在真实世界退化图像时，实际退化与假设不匹配会使得SR方法性能显著下降。为解决退化不匹配时SR方法性能下降问题，提出一种新的盲SR方法，它可以无监督地学习退化图像的退化信息，并使用学习的退化信息有效地调制SR网络。基于对比学习设计了一种无监督的隐式退化表征学习方法，不依赖负样本队列即可提取出准确的退化信息，并提出将经过同一退化的不同图片作为正样本对，克服了图片内容相似性对学习退化的影响，有效提升了隐式退化估计器的表达能力。设计了一个包含全局调制块和通道融合调制块的退化引导调制超分辨率网络，它和退化信息结合可以灵活处理各种退化，以实现最终的SR恢复。实验在四个标准数据集上与多种代表性方法进行了比较，结果显示在×2、×3和×4三种不同放大倍数情形下，在Urban100数据集上比基线模型（DASR）的PSNR分别提高0.85 dB、0.87 dB和0.32 dB。最终提取的退化表征信息质量和重建出的SR图像比DASR均有较大提升。

关键词: 超分辨率, 对比学习, 退化特征学习, 退化特征调制