Industrial image anomaly detection based on multi Gaussian discriminant model and robust core set

Abstract To address the issue of false positive (FP) detections in image anomaly detection caused by the loss of low-frequency features when dealing with high-dimensional feature distributions, we propose the multi-layer Gaussian discriminant anomaly detection model (MGAD). This model utilizes distance metrics based on multiple normal distributions to perform anomaly detection. By mining multi-layer feature combinations from normal samples and incorporating a Gaussian mixture model strategy for pixel-by-pixel probability density estimation, a weighting mechanism is designed to emphasize the role of low-frequency features in Gaussian space. This approach effectively models data collections that do not follow a single normal distribution as a mixture of several Gaussian distributions, thereby reducing false detections. Additionally, we propose a method for calculating the minimum Mahalanobis distance based on the estimation of the minimum covariance determinant. By identifying a subset with the smallest covariance matrix determinant, this method enhances the robust estimation of the data’s central position and spread, thereby reducing the impact of outliers. On the MVTec-AD dataset, MGAD demonstrates outstanding performance with an anomaly detection area under the receiver operating characteristic curve (AUROC) of 98.8%, the anomaly localization AUROC of 98.2%, and the per-class true negative rate for normal samples of 93.1%. Compared with the state-of-the-art models, MGAD improves the detection accuracy for normal samples by 3.6%, demonstrating the best performance among all models. These results highlight the model’s excellent capability in anomaly recognition and reduction of FPs.