Anomaly detection algorithm based on copying and pasting

Abstract

The invention discloses an anomaly detection algorithm based on copying and pasting. The algorithm comprises the four steps of data preparation, feature extraction and reconstruction, training based on a copying and pasting mode and anomaly scoring. The invention provides a novel training mode of an original image, copied and pasted exceptions, a network and a reconstructed image. The original reconstruction function is further upgraded, the original loss function only comprises L2 loss, and the proposed new reconstruction loss function enables the network to compare the similarity between the original image and the reconstructed image from multiple dimensions of brightness, contrast and structure by introducing an SSIM function; the prior art is put forward on the basis of the background that a training set has no abnormal data, which means that the prior art has the defect of lack of prior information, but copying and pasting learning can make up the defect. Firstly, data augmentation can be realized by using a copy-paste learning mode, the defect of scarcity of abnormal data in the prior art is made up, and the common problem of reconstruction error failure in a traditional pixel-level reconstruction model is improved.

Description

technical field [0001] The invention belongs to the technical field of anomaly detection algorithms, and in particular relates to an anomaly detection algorithm based on copy and paste. Background technique [0002] Industrial anomaly detection is often faced with the fact that normal patterns can be clearly defined, but potential anomalous patterns that may exist cannot be defined. Therefore, how to define and distinguish target samples has become a hot spot in academia and industry. Influenced by the field of deep learning, most of the solutions to anomaly detection problems currently focus on deep learning methods, for example: the reconstruction model in the form of an autoencoder, whose purpose is to learn the accurate data distribution of the target sample, and we will A new type of training mode introduces anomaly detection, the purpose of which is to randomly build anomalies to improve the robustness of the model, where copy and paste means random paste, that is, by...

AI Technical Summary

Problems solved by technology

[0003] The original training mode of the industrial anomaly detection algorithm is the original image → network → reconstructed image. The existing technologies are all proposed based on the background of no abnormal data in the training set, which means that the previous technologies will have a lack of prior information. In addition, some training modes with abnormal data training only contain L2 loss, so we propose an anomaly detection algorithm based on copy and paste to solve the problems in the existing technology

Images