XGFu: Enhancing low-light visualization by feature and graph fusion of multiple artificial exposure images

Abstract

Images taken in low-light conditions often have a low dynamic range and include noise; however, existing multi-exposure image fusion methods are frequently affected by color and exposure levels, further complicating the saturation and dynamic range for high-quality images. In addressing these challenges, this paper introduces a two-dimensional graph convolutional multi-exposure image fusion framework (XGFu). It incorporates spatial and channel graph feature fusion for the feature fusion of artificially generated multi-exposure images. Specifically, the proposed enhancement framework consists of two networks: the Multi-Exposure Generation Network (MEG-Net) and the Graph-based Feature Fusion Network (GFF-Net). The MEG-Net combines nonlinear factor estimation modules with weighting schemes to generate a series of artificial exposure images, significantly expanding the hidden feature space of input low-light images. The GFF-Net supports both single-feature set depth fusion and multi-feature set breadth fusion of exposure-related feature sets, with both employing graph convolution to construct a joint reasoning chain of different feature graphs in spatial and channel dimensions. Qualitative and quantitative evaluations conducted on synthetic and real low-light images demonstrate that our model outperforms other state-of-the-art (SoTA) methods in robust low-light enhancement.