Research On A New AnoGAN Algorithm For Industrial Anomaly Detection

With the continuous development of the industrial age,my country’s manufacturing industry is undergoing an era of digital transformation,and intelligent industry has become the focus of attention of various manufacturing companies and enterprises.As one of the research directions of intelligent industry,the field of image anomaly detection is attracting more and more attention.Due to the vigorous development of computer and deep learning technology,the performance of traditional image anomaly detection methods is not as good as that of deep learning-based anomaly detection methods.How to improve efficiency and reduce labor costs as much as possible is gradually becoming a problem in the manufacturing industry.matter of great concern.However,most of the current anomaly detection methods based on deep learning belong to supervised learning methods.Supervised learning methods often require a large amount of labeled data and large labor costs,and the efficiency is low.In the process of practical application,there are Certain limitations.Therefore,this paper proposes a new AnoGAN unsupervised anomaly detection network for industrial anomaly detection,and performs anomaly detection on industrial anomaly data sets.The research content mainly includes the following two points:(1)A new unsupervised anomaly detection network is proposed.Abnormal image data usually has problems such as unbalanced number of samples and unknown abnormal structure,so this paper proposes an unsupervised learning algorithm to detect and locate abnormal images.This method does not need to use marked data,solves the problem of rare image data and unbalanced samples,reduces labor costs,and has a high accuracy rate.(2)A new method for anomaly generation is proposed.Generating high-quality anomaly images is a key factor in improving the performance of anomaly detection models.The traditional Perlin noise generator generates global random noise,which causes the generated anomalies to easily appear in the background of the image,which may be unfavorable for model learning.of samples.In order to solve this problem,this paper proposes a novel anomaly generation module combined with the target detection algorithm LDF(Label Decoupling Framework),which aims to generate anomalies in more prominent areas of the image rather than background areas,thereby generating high-quality anomalies sample.