基于退化感知的盲超分辨率模型

Blind super-resolution model based on degradation-aware

现有的超分辨率方法假设了从高分辨率图像到低分辨率图像的预定义退化过程,这对于具有复杂退化类型的真实世界图像来说很难成立。针对该问题,本文提出了基于退化感知的盲超分辨率模型。该模型用随机模糊核生成低分辨率图像,用对比学习退化表征。模型生成器由包含多个退化感知块的残差组构成。退化感知块用退化表征和图像特征做交叉注意力计算空间权重图,此外模型还收集残差组输出的层级特征,计算层间注意力来复用层级特征,使得模型更加关注高频细节,模型特征提取能力进一步提高。通过消融实验验证了各模块的有效性。在多个国际公开测试集上,放大倍数为4的平均PSNR提升1.45 dB,SSIM提升0.058。实验结果表明,该模型在盲超分辨率任务上取得了显著的性能,且具有良好的视觉效果。

Existing super-resolution methods assume a predefined degradation process from high-resolution images to low-resolution images,which is difficult to hold for real-world images with complex degradation types.For this problem,a blind super-resolution model based on degradation-aware is proposed.The model generates low-resolution images with random blur kernels and learns degenerate representations with contrasts.The model generator consists of residual groups containing multiple degradation-aware blocks.Degraded perceptual blocks use degraded representations and image features to do cross-attention to calculate spatial weight maps.In addition,the model collects layer-level features from the output of residual groups and calculates inter-layer attention to reuse layer-level features.This enables the model to pay more attention to high-frequency details,and the model feature extraction capability is further improved.The effectiveness of each module is verified by ablation experiments.On multiple international public test sets,the average PSNR with a magnification of 4 is increased by 1.45 dB,and the SSIM is increased by 0.058.Experimental results show that the model achieves significant performance on blind super-resolution tasks with good visual results.