双回归网络的单图像超分辨率重建

Dual Regression Networks for Single Image Super-Resolution

现有基于卷积神经网络的单图像超分辨率模型存在三个限制。理论上存在无限的HR图像,可以下采样到相同的LR图像,可能的函数空间非常大。因为现实世界潜在的下采样方法是未知的,使用特定方法配对的数据训练的模型在实际应用中泛化能力差,产生适应性问题。忽视残差分支的高频层次特征。针对上述问题,提出双重回归方案。除了学习从LR到HR图像的原始回归映射之外,额外学习一个对偶回归映射来估计下采样核并重建LR图像,形成一个闭环提供额外的监督,并在残差结构上引入了傅里叶变换,增强模型对高频信息的表达能力。相比其他先进模型以更少的参数重建HR图像,且拥有丰富的高频纹理细节。

There are three limitations to the existing single-image super-resolution model based on convolutional neural networks. Theoretically, there is an infinite HR image that can be downsampled to the same LR image, and the possible function space is very large. Because the potential downsampling method of the real world is unknown, the model of data training paired with a specific method has poor generalization ability in practical application, which causes adaptability problems. The high-frequency hierarchical characteristics of residual branches are ignored. In view of the above problems, a dual regression scheme is proposed. In addition to learn the original regression mapping from LR to HR images, an additional pair regression map is learned to estimate the subsampling core and reconstruct the LR image, forming a closed loop to provide additional supervision, and introducing Fourier transformations on the residual structure to enhance the ability of model to express high-frequency information. Compared with other advanced models, HR images are reconstructed with fewer parameters and rich in high-frequency texture details.