A New Low-light Image Enhancement Based on Disentanglement Representation Learning

Visible light image is the most important information source and has been widely used in many computer vision tasks. However, the existing low-light image enhancement methods focus on generating visual pleasing results without take the downstream CV tasks into account. In this paper, we propose a new low-light image enhancement method based on disentanglement representation learning. Based on an auto-encoder network, we decompose a low-light image into a content feature and an illumination feature at first; then a detail enhancement module is designed to enhance the content feature. At the same time, rather than adjusting illumination feature directly, a downstream CV task oriented illumination enhancement strategy is devised. A preferred illumination alternative can be selected from a set of pre-extracted illumination feature adaptively by the subsequent vision tasks. At final, a decoder reconstructs the enhancement image. Our method tries to collaborate with the downstream CV tasks to pursuit the best whole performance. Quantitative and qualitative experiments demonstrate the advantages of the proposed method.