Low light image enhancement based on multi-branch fusion attention mechanism

A decomposition task-based multi-branch low-light enhancement network (MANet) is proposed for low-light images with poor quality due to overexposure caused by low contrast, loss of detail information, color distortion and noise artifacts, etc., by fusing multi-branch and attention mechanisms. Firstly, the low-light input image is subjected to VGG-19 convolution for feature acquisition at the same time in order to reduce the computational effort, the normal convolution is replaced by a network with separable convolution to extract effective features such as edges and textures. Then in order to be able to adaptively perform adaptive brightness enhancement and noise artifact suppression for different regions in the image, an attention mechanism is introduced in the enhancement module, which is used to learn and set different weight information to obtain feature information for enhancement. Finally, in order to further improve the image quality, multi-scale feature fusion is used in the fusion module, which enables further fusion and enhancement of contextual information. The experimental results show that MANet can adaptively enhance image brightness while reducing image noise and removing artifacts, and improves PSNR by 7% and SSIM by 3.3% compared with GLADNet network.