基于递归多尺度卷积网络的图像超分辨率重建

Image Super-Resolution Reconstruction Based on Recursive Multi-scale Convolutional Networks

不断加深网络的深度可提高网络的超分辨率重建效果,但是网络的加深会导致网络参数量急速增加,难以进行网络训练和内存存储.为了减小深度网络的参数规模并尽量保持网络的重建性能,基于递归和多尺度的思想,文中提出精简的基于递归多尺度卷积网络的图像超分辨率重建方法.首先利用多尺度模块充分提取图像在不同尺度下的特征信息,再通过递归操作实现网络规模的加深而不增加网络的参数量,最后将每次递归操作的输出进行特征融合,作为高分辨率图像重建的输入.实验表明,文中方法在网络参数量较少时重建效果较优.

The performance of image super-resolution reconstruction networks is improved by deepening the depth.However,deepening the network makes the number of parameters increase rapidly,and thus it is hard to train the network and store the memory.To reduce the scale of the deep network and keep its reconstruction performance as much as possible,a concise recursive multi-scale convolutional network is proposed for super-resolution reconstruction based on the concepts of recursion and multi-scale.Firstly,the multi-scale module is employed to extract the features of the image with different scales.Then,the network is deepened by the recursive operation without increasing the number of network parameters.Finally,the outputs of each recursive operation are fused as the input for the reconstruction part.Experimental results show that the network parameters of the proposed method are fewer than those of some existing super-resolution methods with better reconstruction results.