A Research Of Unsupervised Image Anomaly Detection Method With Deep Learning

Image anomaly detection has a wide range of application needs in scenarios such as autonomous driving,intelligent monitoring,and intelligent industry.Although deep learning-based image anomaly detection algorithms have achieved good detection performance,there is still a large room for improvement in unsupervised image anomaly detection without label information,making it one of the most challenging tasks in the current field of computer vision.Unsupervised image anomaly detection involves two tasks: one-class image anomaly detection and multi-class image anomaly detection.One-class image anomaly detection only includes one type of data during the training phase,which greatly limits the model’s representational capacity.Although multi-class image anomaly detection includes multiple types of data,the constraint of having no label information only allows it to explore instance-level representations from a data perspective.In response to the limitations of the two tasks mentioned above,this article constructs frameworks that can effectively improve the detection performance.The specific work is as follows:(1)To address the problem of the lack of suitable training optimization objectives in one-class image anomaly detection methods,which leads to pattern collapse and the loss of representation capacity,this paper proposes a locally optimized pretrained feature fine-tuning framework.By analyzing the root cause of pattern collapse,this framework introduces a local optimization center and designs a K-Normal Nearest Neighbors module,an adaptive projection module and a normal self-attention module,which can improve the representational capacity of fine-tuned features.Additionally,it also introduces a constraint term to ensure the effectiveness in the training phase.Experiments on multiple standard datasets for one-class image anomaly detection demonstrate that the proposed framework outperforms other representative methods and effectively improves the performance of one-class image anomaly detection.Using the AUROC metric,the proposed framework achieves 97.0%,96.5%,89.9% on the CIFAR10,CIFAR100,and MVTec datasets,respectively.(2)Aimting at the current situation that the existing multi-class image anomaly detection works only focus on mining instance-level representation from the data perspective but ignore the semantic information,this paper proposes a cluster-aware contrastive learning framework.Based on the traditional contrastive learning framework,this framework uses clustering algorithms to mine the semantic centers of normal samples and designs a semantic-level contrastive loss function that includes semantic center contrastive loss and pseudo-supervised contrastive loss to fully explore semantic-level features.In addition,this paper compares and analyzes the effects of introducing different layer features,iterative updating of cluster centers,and pre-training on the model’s performance to ensure the effective training of the algorithm framework model.Verification experiments on multiple standard datasets demonstrate that the proposed cluster-aware contrastive learning framework outperforms other contrastive learning strategies.Using the AUROC metric,this proposed framework achieves 97.3%,95.9% on the LSUN(Fix)and Image Net(Fix)datasets,respectively.