Multi-Branch Interaction Network for Low-light Image Enhancement

Abstract

True color information and texture details are challenging to capture while taking pictures in low light. Taking that into consideration, we propose a multi-branch interaction network (MBIN). Specifically, our method combines the theory of CNN and transformer to obtain abundant color information with two branches respectively. Additionally, a channel attention module is designed to obtain more important details and address the issue of gradient fragmentation and gradient dispersion. Ultimately, the final output images are generated after a fusion module which takes the results of the mentioned two branches into consideration. Compared to other state-of-the-art methods, extensive experiments show the result that MBIN performs better than them when it comes to obtain color information and restore the real-world images.