Real-Time Low-Light Image Enhancement Method for Train Driving Scene Based on Improved Zero-DCE

Abstract

Low-light train driving scenes often suffer from issues such as poor visibility, low contrast, and image noise, which seriously affects the accuracy and safety of train detection. In this paper, we propose a real-time low-light image enhancement method based on an improved Zero-DCE algorithm. We first establish a mathematical model of the light curve based on a high-order polynomial function. Then we use the encoder-decoder architecture to build the parameter estimation network for the illumination curve. To ensure real-time performance, the lightweight MobileNet v3 is utilized as the encoder, and the LR-ASPP module is employed for feature fusion and decoding to guarantee real-time performance. Finally, we design a set of self-supervised losses for model training without reference labels. Our experiments on a self-built train driving dataset show that the proposed method outperforms the original images in terms of image details, color vividness, overall evaluation score, and inference time of only 2.84 ms. Our method has important practical and theoretical significance and is of great value for application in railway transportation.