Research On Multi-Case Anomaly Detection Algorithm Based On Feature Information And Memory Bank

Defect detection algorithm is one of the important and challenging research areas in industrial non-destructive testing,which provides efficient guarantees for industrial production safety.Currently,mainstream defect detection algorithms can be divided into two categories: supervised and unsupervised.The former relies on training samples containing defects to identify defects.However,there may be missed detections when defects appear in the testing samples that are not present in the training dataset.The latter reflects all regions that differ from the background by training background data without defects,thereby ensuring the generalization of defect detection,but false positives and missed detections still exist.Although subsequent research has made some improvements on these two methods and achieved good results,they still face problems such as overfitting.In light of these limitations,this thesis conducts in-depth research by taking semi-supervised training mode,dynamic memory bank,and metric learning as entry points to achieve high-precision and low false positive defect segmentation in multiple scenarios in a small number of iterations.The main contributions of this thesis are as follows:(1)This paper proposes a semi-supervised anomaly detection algorithm that combines a metric learning module and a memory bank.This thesis conducts in-depth analysis on the problems of missing detections and false positives from the loss function and feature space perspectives,and proposes two improvement ideas.First,the variable sampling-based improvement selects appropriate variables using the K-Means clustering algorithm and Triplet Loss to adjust the distribution of the feature space.Second,the loss function-based improvement balances the deviation of the loss function value by dynamically assigning strong attention mechanisms to small defect samples using dynamic weight parameters.The experimental results demonstrate the effectiveness of the proposed module,especially in improving defect segmentation accuracy and antiinterference robustness.(2)To improve the algorithm’s generalization ability in different cases,this thesis proposes a multi-case memory bank and parameter inheritance module to achieve rapid generalization in unknown test scenarios with minimal iterations.The proposed multicase memory bank and parameter inheritance module effectively utilize the information in different training cases,enabling it to have certain prior knowledge before the training step begins.This module significantly reduces the number of iterations required for the algorithm to be retrained in new scenarios,improving the efficiency of detection tasks.(3)This thesis proposes a dynamic updating mechanism for memory bank.The mechanism combines multiple-case memory banks and inheritance modules to effectively enhance the representation of the memory banks in different cases and avoid the model falling into local optima.In addition,this thesis proposes a Bank-Case matching module based on KL divergence and image gray value distribution to improve the accuracy of anomaly calculate.This thesis has been validated on test datasets and ablation experiments,further improving the detection ability and anti-interference robustness of the algorithm.