FLA-Net: multi-stage modular network for low-light image enhancement

Nana Yu, Jinjiang Li, Zhen Hua

Abstract: Under the condition of low illumination, the image is easy to appear unclear, the contrast is not enough, and the details can not be fully displayed, which will inevitably bring huge obstacles to the computer vision task. However, low-light image enhancement is a challenging task. Therefore, we design a multi-stage modular network FLA-Net (F: Feature Aggregation Module, L: LBP Module, A: Attention Module) for low-light image enhancement. In this study, we divide the task into three stages: the first stage is the feature extraction stage (FE stage), in which LBP module is added to the feature extraction stage to help the network better recover the image texture details and other information; the second stage is the feature aggregation stage (FA stage), in which the feature aggregation module is added, which can help the network integrate deeper contrast information and color information; the third stage is the image enhancement stage (IE stage), in which the channel attention module is added to the image enhancement stage to make the network pay more attention to the dark areas of the image, which is conducive to the enhancement and recovery of low-light images. Our FLA-Net method solves a series of problems existing in the existing low-light enhancement methods. In addition, we have verified the effectiveness of FLA-Net on multiple public datasets and compared with multiple existing methods for image enhancement. A large number of experimental results show that the enhanced image generated by our method has better subjective visual quality and is better than the most advanced low-light enhancement method in several objective evaluation indicators.