Research on image anomaly detection technique for pipeline leakage
Jikun Guo, Canpeng Zheng · 2024
With the increasing demand of the whole society for oil and gas resources, the safety and reliability of pipelines cannot be ignored, and the abnormal detection of pipeline leakage becomes especially important. The traditional manual inspection method is not only time-consuming and labor-intensive, but also susceptible to the influence of subjective factors, and unable to realize efficient and accurate anomaly detection. Therefore, with the rapid advancement of artificial intelligence for the society, this paper is dedicated to the development of advanced computer vision techniques to automate the discovery of abnormal regions from large-scale oil and gas pipeline image data. In this paper, we propose a semi-supervised anomaly detection method based on teacher-student network, which is applied to the detection of abnormal regions in oil and gas pipeline images. The key of the method is to use the student network to learn the normal patterns from a small amount of labeled data, and at the same time, through the mutual guidance with the teacher-student network, we can fully explore the potential knowledge of a large amount of unlabeled data, and further optimize the performance of anomaly detection. We have specially designed the network structure and loss function for the characteristics of oil and gas pipeline images in order to improve the extraction ability of image anomaly features.