Abnormal image detection method based on self-attention generative adversarial network

Abstract

The invention relates to the technical field of industrial automation, in particular to an abnormal image detection method based on a self-attention generative adversarial network, and the method comprises the following specific steps: S1, obtaining a to-be-detected image; S2, inputting an image to be detected into the TransGANormaly, and obtaining an abnormal score; and S3, judging whether the abnormal score is greater than a certain specific threshold value or not. According to the method, a self-attention mechanism is designed and used for replacing convolution operation, so that the defect that the convolution operation can only extract local features can be effectively overcome, feature extraction on a larger scale level is realized, the anomaly detection performance of the model is effectively improved, the problems that normal and abnormal samples need to be used at the same time at present, the application range is limited, a convolutional neural network is used for carrying out feature extraction and coding on an image at present, convolution operation focuses on extraction of local information, grasp of global information is greatly limited, the scale of an abnormal area in an industrial image may be large, and a method based on convolution operation cannot be well applied.

Description

technical field [0001] The invention relates to the technical field of industrial automation, in particular to an abnormal image detection method based on self-attention generation confrontation network. Background technique [0002] Anomaly image detection is an important problem in the field of industrial automation. Considering that in most cases, abnormal samples are extremely scarce, therefore, the semi-supervised method that only uses normal samples to build a detection model has strong practical application value, and the abnormal detection method based on Generative Adversarial Network (GAN) can Inferring whether the image is abnormal by comparing the feature difference between the original image and the reconstructed image in the latent space, so as to improve its anti-interference ability against noise, and has strong robustness. The existing GAN-based method has cited a patent as a An abnormal image detection method based on a supervised generative adversarial ne...

AI Technical Summary

Problems solved by technology

Considering that in most cases, abnormal samples are extremely scarce, therefore, the semi-supervised method that only uses normal samples to build a detection model has strong practical application value, and the abnormal detection method based on Generative Adversarial Network (GAN) can Inferring whether the image is abnormal by comparing the feature difference between the original image and the reconstructed image in the latent space, so as to improve its anti-interference ability against noise, and has strong robustness. The existing GAN-based method has cited a patent as a An abnormal image detection method based on a supervised generative adversarial network, the patent application number is 201811368737.3, but it has two obvious defects: 1. This method needs to use normal and abnormal samples at the same time, which limits its application scope; 2. This method uses The convolutional neural network extracts and encodes the features of the image. The convolution operation focuses on the extraction of local information, which has great limitations in grasping the global information, and the scale of the abnormal area in the industrial image may be relatively large. Based on the convolution operation method does not work well for

Images