Unsupervised anomaly detection method, system and device and storage medium

Abstract

The invention discloses an unsupervised anomaly detection method, system and device and a storage medium, multiple student networks are used for jointly participating in prediction, the influence of randomization when a single student network is used is avoided, and better model robustness is achieved; a feature pyramid structure is used, and multi-scale features are used, so that multi-scale abnormal conditions can be dealt with; the semantic segmentation network is used to further assist in improving the prediction result, and the semantic segmentation network is used to assist in correcting the result of the previous step by means of the excellent recognition capability of the abnormal region contour, so that the overall method achieves the best effect. Experiments prove that the method obtains the highest accuracy and can be better applied to an anomaly detection task.

Description

technical field [0001] The present invention relates to the technical field of computer vision detection, and in particular, to an unsupervised abnormality detection method, system, device and storage medium. Background technique [0002] Anomaly detection is a traditional task in computer vision and has important applications in industry, medical, transportation, and other fields. Anomaly detection combined with deep learning, as an important means of industrial automation maintenance and repair, is increasingly applied in various industrial production scenarios to improve maintenance efficiency and ensure personnel safety. However, the effect of commonly used deep learning methods usually relies on a large batch of high-quality training data. In actual production scenarios, the probability of occurrence of abnormal conditions is low and irregular, so it is very difficult to obtain abnormal data. The variety is difficult to predict, and the anomaly detection network model ...

AI Technical Summary

Problems solved by technology

[0005] For the problem of anomaly detection, the above three schemes are unsupervised anomaly detection methods based on single-category classification. The first two schemes use the structure of encoder and decoder, use the reconstruction error to train the network and perform anomaly detection, but only use For deep-level features, the ability to detect abnormalities of multiple scales at the same time will be greatly reduced. At the same time, the results generated based on only one network may have random effects, and the detection results are not robust enough; the third scheme uses historical images to compare real-time images. For anomaly detection, the data requirements are high. It is necessary to ensure that there is no abnormality in the historical image and that it is similar to the real-time image as a whole. The preconditions of the method are strict and it is difficult to migrate. The detection effect of the edge of the abnormal area is very poor, and it is difficult to convert it into a commonly used semantic map for detection

Images