Night Low-Light Image Enhancement Using Variational Retinex Model and Activation Map

Abstract

The illumination enhancement is important for the computer vision applications using outdoor. Although many studies have been reported to solve this problem, they contain various problems. Therefore, we used a method combining variational retinex model and bright channel prior to suppress contrast over-emphasis and noise. We also reduced the computational cost by using FFT as a method to solve the quadratic optimization problem. Furthermore, in order to suppress blur artifacts, we used activation map changed to a form applicable to night images. In our experiments, we confirmed the effectiveness of our method, and improvement of detection accuracy by object detection algorithm.