A Decoder-Free Reconstruction Method for Semi-Supervised Surface Defect Detection

Chen Liu,Zhenyu Shi,Shibo He,Shunpu Tang,Qianqian Yang

DOI: https://doi.org/10.1109/ticps.2024.3456758

2024-01-01

Abstract:Detecting defects on railway tracks is critical for the operation of high-speed trains. Despite a plethora of machine vision-based methods designed to tackle this problem, the majority adopt a supervised setting and demand considerable labeled training data, inclusive of defective samples, which is expensive and impractical. In this paper, we propose an I nvertible R econstruction neural N etwork (IRNet) for semi-supervised rail surface defect detection, where only normal images are accessible during training. Firstly, we devise an information-preserving feature encoder comprising several invertible blocks. This structure safeguards subtle visual patterns distinguishing normal and defective images from being obscured by background information, guaranteed by its mathematical reversibility property. Second, to overcome the overgeneralization issue of conventional autoencoders caused by imperfectly crafted decoders, we propose a novel decoder-free reconstruction workflow based on the invertible feature encoder. Specifically, we force one portion of extracted features to approach a predefined constant tensor during the training stage by minimizing their mean squared error. Next, we feed the remained features and the predefined constant tensor backward into the encoder to reconstruct the original images. During the testing phase, we formulate an anomaly score that consolidates the reconstruction error and mean squared error to spot defects. Extensive experiments are conducted on 4 real-world datasets. Our method consistently outperforms state-of-the-art techniques, demonstrating an average increase of 8.5% on the F1 score.