Research on Data Enhancement Method for Complex Pavement

Abstract

In the era of intelligent vehicle development, environmental awareness technology as the basis of vehicle decision-making and route planning, affected by the weather, accurate road image recognition is particularly important. Aiming at this problem, this paper constructs a large scale pavement image dataset and proposes an improved image generation method based on CycleGAN data enhancement. The core innovation of the method is the improvement of the model structure, introduce Wassertein distance, and implement global Lipschitz constraint with spectral constraint. The experimental results demonstrate that the accuracy of the data set is up to 91.85%, which is 6.41% higher than the original CycleGAN, and the convergence speed is faster.