Improving the Anomaly Detection Performance of a Geometric Transform-based Convolutional Network

Using deep learning (DL) technology, neural networks have achieved great success in various fields of computer vision. Among them, anomaly detection is a promising application of image defect analysis. The purpose of the detector is to find the out-of-distribution when predicting the probability of a DL network for abnormal samples, after some normal sample images are given for training. Geometric transformation (GT) based anomaly detection is one of the recent best methods for classifying abnormal samples among many normal ones. However, the GT method training process is unstable and too inaccurate to be used in industrial applications. The goal of this study is to suggest a method to improve the performance of a GT-based anomaly detector (GTnet). Using observations of GTnet behavior and its training properties, we propose the addition of three techniques that can improve anomaly detection performance for defect inspection in a factory production process. Specifically, k-Winners-Take-All (k-WTA) was applied to the GTnet base model to resist data corruption such as dust on the sample, the temperature scaling method was added to consider correlations between GT classes with similar appearance, and loss redefinition was applied to improve the efficiency of optimal training. Accuracy was improved from 98.56% to 99.86% in the inspection of vehicle part assembly defects, which requires extremely high accuracy. Experimental evaluations were conducted to verify the performance improvement of the GT anomaly detector.