Super-Resolution Image Reconstruction Algorithm Based on Improved Information Distillation Network

Recently, convolution neural networks have been demonstrated remarkable process on single image super-resolution. However, as the depth of the networks increase, the computing time also increase. At the same time, there are higher requirements for the hardware computing power and memory. Information distillation network is a lightweight super-resolution reconstruction network, which has the advantages of fast running speed and few network parameters. However, there are still some deficiencies in the utilization rate of shallow features and the high-frequency details extracted from features. In this paper, an improved information distillation network is proposed. On the one hand, the convolution kernels of different scales are used to extract the high-frequency information of images under different scales, and the feature maps of different scales are fused and put into the enhancement unit to increase the feature channels; on the other hand, the idea of dense connection is used to build a network to improve the utilization rate of shallow features. Experimental results on the Set5, Set14 and BSD100 datasets show that the proposed algorithm improves the Peak Signal-to-Noise Ratio and Structural Similarity compared with the five mainstream algorithms such as Bicubic, SRCNN, VDSR, LapSRN and IDN.