Semantically Contrastive Learning for Low-Light Image Enhancement

Low-light image enhancement (LLE) remains challenging due to the unfavorable prevailing low-contrast and weak-visibility problems of single RGB images. In this paper, we respond intriguing learning-related question -- if leveraging both accessible unpaired over/underexposed images high-level semantic guidance, can improve performance cutting-edge LLE models? Here, propose an effective semantically contrastive learning paradigm for (namely SCL-LLE). Beyond existing wisdom, it casts task as multi-task joint learning, where is converted into three constraints brightness consistency, feature preservation simultaneously ensuring exposure, texture, color consistency. SCL-LLE allows model learn from positives (normal-light)/negatives (over/underexposed), enables interact with scene semantics regularize network, yet interaction knowledge low-level signal prior seldom investigated in previous methods. Training on readily available open data, extensive experiments demonstrate that our method surpasses state-of-the-arts models over six independent cross-scenes datasets. Moreover, SCL-LLE's potential benefit downstream segmentation under extremely dark conditions discussed. Source Code: https://github.com/LingLIx/SCL-LLE.