Unsupervised industrial image ensemble anomaly detection based on object pseudo-anomaly generation and normal image feature combination enhancement

With the development of industrial video technology, the use of cameras rather than a variety of expensive sensors to obtain process or product data has gained more attention. One of the important applications is the use of image data for anomaly detection. It is difficult to collect anomaly data in actual engineering practice, which makes the anomaly detection of industrial products often need to be carried out under the condition of a single data type. How to achieve anomaly detection without anomaly data has become a new challenge. An unsupervised ensemble anomaly detection method based on image enhancement is proposed for image detection with normal data only. The proposed method first uses local pseudo-anomaly generation and object location to generate high-quality pseudo-anomaly images. Then, the pseudo-anomaly images and pseudo-labels are used to guide the training of a reconstruction model and a self-supervised model. In the detection phase, an unsupervised feature screening method is designed to extract sensitive filters, and the normal image features in the feature space output by these sensitive filters are combined and enhanced. Finally, ensemble detection is implemented using different anomaly scores. The experiments show that the proposed method can achieve performance improvements in 15 real datasets.