Research On Anomaly Detection Method Based On Attention Mechanism

The industrial anomaly detection is a crucial part of industrial production and is important to ensure product quality.At present,most factories still use manual or traditional computer vision methods for anomaly detection,which may inevitably suffer from inefficiency,false detection and omission in some complex situations.In recent years,with the development of deep learning technology,many researchers have introduced deep learning into the field of industrial anomaly detection.Compared with traditional computer vision methods,deep neural networks have stronger feature extraction and generalization capabilities,and can adapt to more complex anomaly detection situations and have higher detection accuracy.Therefore,in this paper,we focus on the problems in industrial anomaly detection scenarios and conduct research on anomaly detection methods based on deep learning technology.The specific research contents are as follows:(1)We summarize the current mainstream industrial anomalydetection algorithms,introduce the datasets used in the experiments,and discuss the difficulties and challenges encountered when implementing the mainstream algorithms in industrial anomaly detection scenarios based on the datasets.(2)To address the problem of various types of anomalies in industrial anomaly detection scenarios and the large differences between different types of anomalies,this paper tries to introduce the attention mechanism in the model and proposes a feature fusion module based on the attention mechanism.In this feature fusion module,we try to combine the convolution and attention mechanisms so that the model can obtain the advantages of both.We propose four different ways to combine convolution and attention mechanisms,and we experimentally demonstrate that among them,the feature fusion module with the addition of both multi-head self-attention and shortcut connections has the best performance.Thanks to the stronger fitting ability and generalization ability of the attention mechanism,our model can detect the location of anomalous regions more accurately and has a certain performance improvement in the anomalous region segmentation task.(3)Although the attention mechanism improves the performance of the model,it also makes it more difficult to train the model,and difficult to achieve optimal performance with few training samples.To address the above issues,we add cross-layer connections to the model,inspired by some models in image segmentation.The introduction of skip layer can provide rich detail information to the decoder and rectify the bias generated by the decoder in reconstructing the encoder feature map.Finally,we propose an attention based reverse reconstruction model with 98.9%AUC-ROC in anomalous image classification task,98.1%AUCROC in anomalous region segmentation task,and 94.7%AUC-PRO on MVTec AD dataset,and validating the effectiveness of the method.