Defect Detection of Leakage Cable Fastener in Railway Tunnel Based on Block Reconstruction GAN.

Owing to the fuzzy structure, small background difference, diverse types and complex data distribution, the defect detection of railway communication cable fasteners is still a challenge. Aiming at the problem that the existing CNN methods are sensitive to the distribution of defect data, the unsupervised GAN is introduced, and a block-reconstruction-GAN-based algorithm is proposed. Firstly, the YOLOv5-M model is used to locate the image of the leaky cable fastener. Then, a block-removing approach based on existing Skip-GANomaly is proposed to split the fastener training samples, which can alleviate the overfitting problem of Skip-GANomaly during the reconstruction process of defect fastener by dispersing fastener features and increasing the number of samples. In addition, the cosine similarity is proposed to calculate the similarity score between the original fastener image and reconstructed image, which will be used to distinguish the defect fastener. The experiments are carried out on the cable fastener dataset collected by the comprehensive inspection vehicle and the experimental results show that the AUC value of the proposed algorithm for multi-type fastener defect recognition reaches 0.961, which is 3% higher than that of Skip-GANomaly algorithm. In addition, the visualization results show that compared with other two classical anomaly scores, the method based on cosine similarity can better distinguish defect fastener from the normal fastener. [ABSTRACT FROM AUTHOR]