Research On Anomaly Detection Algorithm Based On Self-Supervised Learning

Anomaly detection has always been an important technical component in the field of computer vision,and in recent years,with the development of deeplearning,the technical route of anomaly detection has been changing in recent years.In actual production,the probability of anomalous traits.The low probability of anomalous traits and the complex and diverse types of anomalies lead to uneven distribution of anomalous samples.To improve this situation,the In this paper,we propose a selfsupervised learning framework that combines convolutional neural networks and data augmentation,with the goal ofreduce the reliance of anomaly detection models on a large number of negative sample data and reduce the detection cost.This paper focuses on the above problem.In this paper,we conduct the following research,and the main results include:(1)Based on Cut Paste data augmentation and incorporating some ideas of Scar,we construct a Cut Paste(Scar)data enhancement method is developed based on the Cut Paste data enhancement method,which incorporates some ideas of Scar.By performing random erasing,random patching,crop,rotate,paste and other operations to simulate object anomalies in real scenes,for dust,physical deformation.It has good ability to simulate abnormal traits such as dust,physical deformation,stains,spots,and defects.Then,based on image similarity combined with CNN Convolutional networks are combined with CNN to design an anomaly classification model,by which the invalid anomalous data are eliminated,reducing the negative impact of naive anomalous data on the anomaly detection model.The anomaly detection model is based on the image similarity and CNN convolutional network.(2)A self-supervised learning-based anomaly detection algorithm is constructed.First,an SVM support vector The anomaly detection model is constructed by the SVM support vector model,and then an anomaly detection model with a similar role is constructed based on the YOLOv5 framework.The anomaly detection model is constructed based on the YOLOv5 framework.The self-supervised dataset obtained by data augmentation is used as the training set for both models,and the unprocessed fully supervised dataset is fed into the model.The training results under both datasets are exported.Finally,the original dataset is fed into another classical self-supervised model,DCGAN,for training.The experimental results show that the.The training results of SVM model and YOLOv5 model are close to each other for both datasets,and the training results of self-supervised dataset on some object types are even slightly better.The results show that the SVM model and the YOLOv5 model have similar training effects on the two datasets,and the self-supervised model even has a slight advantage on some object types,indicating that the self-supervised model has achieved the effect of replacing manual labeling to some extent.In addition,when compared with the classical self-supervised model DCGAN,the overall effect is improved by about In addition,compared with the classical self-supervised model DCGAN,the overall effect is improved by about 1.5 percentage points,which makes a new exploration for the field of anomaly detection.