Semi-supervised pipeline anomaly detection algorithm based on memory items and metric learning

Traditional detection algorithms of pipeline non-destructive testing extract information from a large number of defect samples to ensure the detection performance, but even if an adequate of defect samples are collected, it is difficult to enumerate the possible defect morphology in nature. In this paper, we proposed a new semi-supervised anomaly detection algorithm to solve the existing problems. We consider using the most representative feature vectors generated by feature extractor as memory items to represent background information. In addition, this paper also imports a few defect samples to form a semi-supervised structure in the training stage and introduces a metric learning module to make the memory items have the ability to fully represent the background and enhance robustness. To prove the effectiveness of our algorithm, this paper has verified its performance in micro-size pipeline defects. In the experiment, the high-definition industrial camera was used to scan and record the image sequence from the inner surface of the pipeline sample. The latest anomaly detection algorithms have been used as a platform for objective performance evaluation. The result shows our algorithm is more effective in pipeline defect detection and has strong robustness for anti-interference.