U-TEN: An Unsupervised Two-Branch Enhancement Network for Object Detection Under Complex-Light Condition.

The goal of low-light enhancement is to improve the visual quality of dark regions in an image. However, the existing low-light enhancement methods always failed in nighttime traffic surveillance. The reason comes from complex-light instead of low-light, while traditional methods can hardly treat regions with different light conditions. In this paper, in order to solve above challenge, we propose an unsupervised tow-branch complex-light enhancement network (U-TEN) with graph attention network and generative adversarial network to enhance pixels in different regions according to their light conditions. To solve the problem that enhance weights are supposed to varied with regions, we proposed two-branch attention block. It aims to model two kinds of light information interaction. U-TEN can improve the object detection performance under complex-light condition in nighttime traffic surveillance. The proposed U-TEN is verified on ODLS dataset, and the results show that our network improves the object detection average precision by 3.50% on average, and 4.93% in maximum. These results demonstrate the proposed U-TEN has great potential to be used in vision applications of nighttime traffic surveillance.