Dynamically Activated De-Glaring and Detail- Recovery for Low-Light Image Enhancement Directly on Smart Cameras

Abstract

Low-light conditions often significantly affect the stability of a computer-vision system. Existing studies of unpaired-learning-based low-light image enhancement do not consider glare that occurs during the night, which can lead to significant degradation of image quality. To improve image quality, our study proposes an additional enhancement module that can be applied to existing methods. That is, our proposed “lightweight low-light image de-glaring network” can remove glare from low-light images. We also propose a “low-light image-detail-recovery network” to enhance the boundary details of low-light images after removing glare to further improve image quality. The experimental results show that our proposed approaches can effectively improve low-light image quality. In addition, we propose “dynamically activated de-glaring” to assess the quality of input images first to determine whether de-glaring should be undertaken in order to effectively utilize the computational resources of a smart camera and avoid unnecessary image enhancement. The experimental results show that running time and frames per second can be greatly improved when applied to real-world scenarios.