The image super-resolution algorithm based on deep learning has a good reconstruction effect, and the reconstruction can be further enhanced by using multi-scale features. There are different extraction methods for multi-scale features, and current deep learning-based super-resolution algorithms often use only one method when utilizing multi-scale features. We use an error feedback mechanism with a dense residual mechanism to fuse multi-scale features and propose Feedback Multi-scale Residual Dense Network (FMDN), which uses two different multi-scale features to enhance the reconstruction effect. On the other hand, in the previous multi-scale feature fusion often used the method of concatenating. We design a new error feedback-based feature fusion method, and the experimental results show that it has better results than the common method of concatenating. In addition, we further use the feedback mechanism of recurrent to improve the efficiency of the module, which can use fewer layers to achieve the effect of more layers of the basic model, and take up less space, faster, or make a network with a larger number of layers have better results. Compared with the state-of-the-art method, the proposed method shows promising performance according to qualitative and quantitative evaluation.

基于深度学习的图像超分辨率算法具有良好的重建效果，利用多尺度特征可以进一步增强重建效果。多尺度特征有不同的提取方法，目前基于深度学习的超分辨率算法在利用多尺度特征时往往只使用一种方法。我们使用具有密集残差机制的误差反馈机制来融合多尺度特征，并提出反馈多尺度残差密集网络（FMDN），它使用两种不同的多尺度特征来增强重建效果。另一方面，在以前的多尺度特征融合中经常使用concatenating的方法。我们设计了一种新的基于错误反馈的特征融合方法，并且实验结果表明它比普通的拼接方法有更好的效果。另外我们进一步利用recurrent的反馈机制来提高模块的效率，可以用更少的层数达到基础模型更多层数的效果，而且占用空间更少，速度更快，或者做一个网络层数越大效果越好。与最先进的方法相比，所提出的方法根据定性和定量评估显示出良好的性能。