An Industrial Anomaly Detection Method Based on Spatial Neighborhood Kernel Density Estimation and Feature Sparse Representation

In recent years, there has been significant development in computer vision technology, which has been implemented in the field of industrial quality inspection with promising outcomes. Industrial image anomaly detection, a crucial research area in computer vision, aims to identify abnormal data that deviates from the expected normal pattern, ensuring the reliable operation of various systems. Therefore, the application of anomaly detection techniques in industrial image anomaly detection and localization has immense significance. This paper presents an unsupervised image anomaly detection algorithm that leverages the spatial and neighboring context information of the image. It employs the kernel density estimation method for non-parametric modeling to estimate the image's normal distribution. Additionally, this paper introduces a sparse representation-based method that employs existing deep network features to reconstruct the image's depth features through sparse expression. By calculating the error in feature computation for abnormal samples during sparse representation, precise pixel-level anomaly scores can be predicted. The proposed method exhibits excellent performance in experiments, achieving an AUROC score of 99.3% in the field of anomaly detection and localization on commonly used benchmark datasets.