Image Enhancement for Railway Inspections Based on CycleGAN under the Retinex Theory

Abstract

Under the “safe, green, efficient and intelligent” requirements of modern railway construction, it s important and challenging to improve the automated and intelligent railway inspection system, which is an essential part of building domestic transportation infrastructure. Furthermore, the enhancement of images collected in low-light environments is an indispensable prerequisite for the railway inspection system to carry out intelligent analysis. Therefore, a low-light image enhancement means inspired by the Retinex theory based on CycleGAN is advanced. Specifically, a residual neural network is designed to decompose input images. Meanwhile, the unsupervised learning model CycleGAN is selected to adjust brightness maps. In addition, the BM3D algorithm is used to reduce the noise in reflection images. Eventually, extensive experiments on multiple data sets show that the proposed method can better restore image details, obtain more comfortable visual effects and more natural colors. Moreover, the enhancement results are better than the majority of classic methods.