Progressive Dual-Branch Network for Low-Light Image Enhancement

Images captured in low light and backlit conditions are characterized by low brightness, low contrast, and varying degrees of degradation. Simply enhancing image contrast will fully expose hidden noise and color distortion, affecting people's subjective visual perception and performance in other application scenarios. In order to improve the loss of details, color imbalance and artifacts in the enhanced images, we propose a progressive dual branch network(PDBNet) for low-light image enhancement. In this paper, an assisted recovery module(ARM) is designed by exploiting the hybrid correlation and feature complementarity between the inverted image and the low-light image. Feature information at different scales is progressively extracted by cascading multiple ARMs. Considering the network execution efficiency and the amount of parameters, we use depthwise separable convolution(DSC) and asymmetric assisted recovery module(ARM) to improve the computational efficiency of the model. To reduce the degradation caused by enhancing image contrast, the introduction of the large kernel attention(LKA) block allows the network to emphasize hidden low-light information regions, effectively suppressing noise and improving color imbalance. In order to effectively fuse the feature information between the inverse image and the low-light image, an attention fusion block(AFB) is designed. AFB can effectively acquire global feature information and re-encode semantic dependencies between channels. Finally, a fusion reconstruction module(FRM) is designed to further refine the feature information and enhance the information flow between networks. After sufficient qualitative and quantitative experiments in publicly available low-light image datasets, it is known that our method has better visual quality and metric evaluation scores than other state-of-the-art low-light image enhancement methods.