Attention-based Deep Pyramidal Network for Low-light Image Enhancement

Abstract

Various images taken under complex lighting conditions often suffer from degraded image quality. Such poor quality not only fails to meet user expectations but also leads to significant performance degradation in many applications. In this paper, we propose a new low-light image enhancement method that exploits the feature validity of the attention mechanism and the spatial validity of the pyramidal structure. Specifically, the proposed method is able to recover image details from the original input and enhance the lighting and combine them at the end of the network. Moreover, the pyramid structure defined in the feature space based on the rich connection of higher-order residuals in the multi-scale structure makes the recovery process more efficient. This decomposition-based scheme is quite ideal for learning the highly nonlinear relationship between degraded images and their enhancement results. Experimental results on different datasets show that the proposed ADPN outperforms existing methods. The code and model are publicly available at: https://github.com/zhangxueshi0717/Attention-based-Deep-Pyramid-Network-for-Low-Light-Image-Enhancement.