Super-Resolution Reconstruction of Human Face Image Based on Network of Residual Dense Connection Channel Attentional

Abstract

Convolutional neural network is very important for super-resolution reconstruction of face images. However, deeper convolutional networks are more difficult to train for face super-resolution reconstruction. The input low-resolution face images contain a wealth of low-frequency information, these low-frequency information is processed equally in the channel, which hinders the characterization ability of the convolutional neural network, sequentially the recovery performance is affected, so that the super-resolution reconstruction results obtained are inaccurate. In order to solve these problems, we propose a residual convolution network with a compact structure. Specifically, we propose a modified residual structure group to deepen the network structure. In the residual structure group, the compact and dense connection structure is combined to make full use of the high frequency information learned by the previous network, and the features of the shallow small perception field are transmitted to the network to complete the specific task. In addition, we introduce an attentional mechanism for adaptive re-estimation of channel characteristics by considering the dependencies between channels. Experiment results show that the proposed network achieves better accuracy and satisfactory visual effect.