JOF-Net: A Joint Optimization Framework for Low-Light Image Enhancement

Low-light image enhancement (LLIE) is still an essential task in computer vision because of the coupling effect of under-exposure and noise, resulting in severe image quality degradation and performance drop of downstream tasks. Current methods usually handle exposure correction and denoising as disjoint or sequential steps, resulting in noise amplification or loss of detail. In this paper, we present a new Joint Optimization Framework, called JOF-Net, that unifies exposure compensation and noise suppression for low-light image enhancement in a joint architecture. The JOF-Net uses a shared feature encoder and two dedicated modules, Dynamic Exposure Adjustment Module (DEAM) and Noise-aware Attention Denoising Module (NADM). By sharing representations and guiding each other, these modules allow adaptive and content-aware sphere enhancement. Comprehensive experiments on various benchmark datasets show that JOF-Net significantly outperforms existing state-of-the-art LLIE methods in terms of no-reference image quality metrics as well as visual perceptual quality. In addition to simultaneous illumination enhancement and noise suppression, such structural fidelity enhancement makes our method particularly effective in real-world night vision applications.