A Semi-Supervised Railway Foreign Object Detection Method Based on GAN

Abstract

The rapid development of deep learning provides new technical means for railway foreign object detection. However, in practical applications, the datasets of railways with foreign objects are scarce. In order to solve this problem, by improving the loss function and anomaly image evaluation standard, this paper proposes a new semi-supervised anomaly detection method based on GAN (Generative Adversarial Networks). Experiments show that our method can achieve railway foreign object detection without anomaly prior knowledge. Regarding anomaly recognition, a 0.058 AUC (Area Under Curve) and a 6% classification accuracy relative improvement for the railway dataset used in this paper are obtained.