An Image Denoising and Enhancement Approach for Dynamic Low-light Environment

Abstract

Image enhancement technology can greatly improve the visual recognition accuracy of robots in low-light environments. In the past few years, there have been several effective low light enhancement algorithms. However, a supervised deep-learning algorithm requires paired data, which is difficult to collect in dynamic scenes. Meanwhile, there are problems of color distortion and noise amplification in the enhanced image. In this paper, we train an effective image denoise and enhancement model. Furthermore, our method only uses low-light images as training data without ground truth to solve the difficulty of data collection. Extensive experiments on a variety of benchmarks have demonstrated that proposed model is qualitatively and quantitatively better than state-of-the-art methods.