基于梯度感知的单幅图像超分辨

Gradient-aware based single image super-resolution

随着生成对抗网络在图像超分辨(SR)领域的应用,一些感知驱动的SR方法可以恢复出纹理细节更加丰富的SR图像,有效地缓解了由PSNR主导的SR方法导致重建图像趋于平滑的问题。然而梯度信息作为图像纹理的一种重要表现形式,鲜有SR方法能准确、高效地利用。为此,提出一种基于梯度感知的图像超分辨(GASR)算法,可以更准确地利用梯度信息。一方面,使用梯度域的特征图作为作用在图像域特征图上的卷积核,有效地避免了不同域特征图串联所带来的域冲突问题;另一方面,通过对卷积核尺寸等设计细节的调整使两个分支对应位置所输出的图像域与梯度域特征图的感受野一致。此外,由于实际应用对网络轻量化需求的提高,提出的GASR算法还有效降低了参数量和计算量。与同样利用梯度信息的SPSR相比,GASR最终以约1/6的参数量与1/10的计算量取得了与其相近的性能。在Set14数据集上,LPIPS与PSNR分别提升了0.0022与0.217。实验结果验证了GASR可以在纹理生成与图像平滑度之间取得一个很好的平衡,视觉化效果也验证了GASR不仅可以准确地恢复SR图像,而且在一定程度上缓解了杂乱纹理的生成。

With the application of generative adversarial networks in the field of image super-resolution(SR),some perception-driven SR methods can recover SR images with richer texture details,effectively alleviating the over-smoothing problem of the PSNR dominated SR methods.Gradient information is an important representation of image texture.However,few SR methods can make use of this information accurately and efficiently.In this paper,a gradient-aware single image super-resolution(GASR)is proposed,using gradient information better from two aspects.On the one hand,the feature map of gradient domain is used as the convolution kernel imposing on the feature map of image domain,which can effectively avoid the domain conflict caused by the concatenation of feature map of different domains.On the other hand,by elaborating the network details such as convolution kernel size,etc.,the image fields output at the corresponding positions of the two branches are consistent with the feel fields of the feature maps in the gradient domain.In addition,the proposed GASR algorithm also effectively reduces the number of parameters and the amount of computation due to the increased demand for network lightweight in practical applications.Compared to SPSR,GASR can achieve the same performance at the cost of about1/6 of the parameters and 1/10 of the computation of SPSR.On Set14 dataset,LPIPS and PSNR increase by 0.0022 and 0.217,respectively.The experimental results show that GASR can achieve a good trade-off between texture details and image smoothness.In addition,GASR can not only reconstruct high fidelity SR image,but also alleviate the generation of messy textures.