Low-light image enhancement via illumination optimization and color correction

Abstract

The issue of low-light image enhancement is investigated in this paper. Specifically, a trainable low-light image enhancer based on illumination optimization and color correction, called LLOCNet, is proposed to enhance the visibility of such low-light image. First, an illumination correction network is designed, leveraging residual and encoding-decoding structure, to correct the illumination information of the $V$-channel for lighting up the low-light image. After that, the illumination difference map is derived by difference between before and after luminance correction. Furthermore, an illumination-guided color correction network based on illumination-guided multi-head attention is developed to fine-tune the $HS$ color channels. Finally, a feature fusion block with asymmetric parallel convolution operation is adopted to reconcile these enhanced features to obtain the desired high-quality image. Both qualitative and quantitative experimental results show that the proposed network favorably performs against other state-of-the-art low-light enhancement methods on both real-world and synthetic low-light image dataset.