Image Reconstruction Based on Progressive Multistage Distillation Convolution Neural Network

Abstract

To address the problem that some current algorithms suffer from the loss of some important features due to rough feature distillation and the loss of key information in some channels due to compressed channel attention in the network, we propose a progressive multistage distillation network that gradually refines the features in stages to obtain the maximum amount of key feature information in them. In addition, to maximize the network performance, we propose a weight-sharing information lossless attention block to enhance the channel characteristics through a weight-sharing auxiliary path and, at the same time, use convolution layers to model the interchannel dependencies without compression, effectively avoiding the previous problem of information loss in channel attention. Extensive experiments on several benchmark data sets show that the algorithm in this paper achieves a good balance between network performance, the number of parameters, and computational complexity and achieves highly competitive performance in both objective metrics and subjective vision, which indicates the advantages of this paper’s algorithm for image reconstruction. It can be seen that this gradual feature distillation from coarse to fine is effective in improving network performance. Our code is available at the following link: https://github.com/Cai631/PMDN.

Lightweight spatial-channel adaptive coordination of multilevel refinement enhancement network for image reconstruction

“…However, these algorithms perform feature compression brutally, which can result in the loss of some key information. To this end, Cai [26] improved the distillation method, gradually distilling the key features of each stage from coarse to fine, and then using convolution layer to polymerize these features. Wang [13] proposed a scheme for adaptive learning of airspace and channel masks, which greatly reduces the computational complexity of the network while maintaining the performance of the network.…”

“…Niu [33] designed a layer attention block and a channel space attention block to more comprehensively and selectively exploit information-rich features by modeling the inter-dependencies between different layers, channels, and locations. Cai [26] proposed a weight-sharing information lossless attention block, which enhances the recovery of high-frequency information such as edge textures through a weight-sharing auxiliary branching module. Wang [34] constructed a lightweight recurrent residual channel attention block to further improve the network performance by introducing recurrent connections in the attention module with a reduced number of module parameters.…”

“…In addition, we use splicing and convolution layers to aggregate the important features at each stage. The WSILB-V used here is a variation of the WSILB(Weight-Sharing Information Lossless Attention Block) proposed in the paper [26], with the change that we use two 1×1 convolution layers instead of one 1×1 convolution layer to achieve feature compression and recovery, making it more lightweight.…”