AUTHOR=Liu Fangjin , Hua Zhen , Li Jinjiang , Fan Linwei 

TITLE=Low-Light Image Enhancement Network Based on Recursive Network

JOURNAL=Frontiers in Neurorobotics

VOLUME=Volume 16 - 2022

YEAR=2022

URL=https://www.frontiersin.org/journals/neurorobotics/articles/10.3389/fnbot.2022.836551

DOI=10.3389/fnbot.2022.836551

ISSN=1662-5218

ABSTRACT=In low light environment,the light source obtained by the image acquisition device is not sufficient, pose a huge obstacle to other computer vision tasks.This paper proposes a multi-scale feature fusion image enhancement network based on recursive structure,using hybrid attention CBAM modules in the network,to obtain multi-scale features,we combine the inception model and the U-Net model to form the MIU module(Multi-scale inception U-Net),the whole network is learned in T recursive stages,the input of each stage is the original low-light image and the intermediate enhancement result of the previous recursive output.In the t-th recursion,we first use the hybrid attention module CBAM to fuse the channel feature information and spatial feature information of the low-light image,make the network more focused on the low-light areas of the image.Next,We use the MIU module to extract multi-scale features,obtain 3 different scales of features,and fuse the multi-scale features to get the intermediate enhanced image results.Finally,the intermediate enhanced image is spliced with the original input image and fed into the (t+1)-th recursive iteration, with the intermediate enhanced result providing higher-order feature information and the original input image providing lower-order feature information, the whole network constitutes a cyclic recursive pattern,and the enhanced result is output after multiple cyclic recursions.We use multiple methods to conduct experiments on multiple public datasets and analyze the results subjectively and objectively.The experimental results show that although the network structure of this paper is simple,but compared to other methods the method in this paper can better recover the details and increase the brightness of the image, reduced image degradation.