A deep thermal-guided approach for effective low-light visible image enhancement

Low-light visible image enhancement is important for various visual computing applications under conditions of poor lighting or hazardous weather. However, existing low-light image enhancement methods are mostly based on a single visible channel and cannot achieve satisfactory performance when processing real-captured nighttime images. In this paper, we attempt to utilize the complementary edge/texture features presented in thermal images to provide a stable guidance map to facilitate the enhancement of features extracted on low-light visible images. For this purpose, we propose a novel Central Difference Convolution-based Multi-Receptive-Field (CDC-MRF) module to effectively extract multi-scale edge/texture features on thermal images. Then, we design a thermal-guided convolutional block (TGCB) to enhance the low-light visible features under the guidance of thermal features. To our best knowledge, the proposed thermal-guided low-light image enhancement network (TGLLE-Net) represents the first attempt to perform low-light visible image enhancement by incorporating complementary information presented in both visible and thermal channels. The advantages of the proposed TGLLE-Net are twofold. Firstly, it is capable of suppressing severe noise disturbance presented in low-light visible images under the guidance of low-frequency components in thermal images. Moreover, TGLLE-Net can promote detail/appearance restoration of objects with distinctive thermal features (e.g., pedestrians, vehicles, and buildings). Both objective and subjective evaluation results demonstrate that our proposed TGLLE-Net outperforms state-of-the-art methods in terms of restoration accuracy, visual perception, and computational efficiency.