Salient Textural Anomaly Proposals and Classification for Metal Surface Anomalies

Abstract

With a drastic growth in the development of Automated Visual Inspection (AVI) systems in the industries, the capabilities of such applications to aid human inspectors for anomaly localization and identification have also increased. However, some issues with anomaly detection and classification in AVI systems are that such anomalies are rare in occurrence and exhibit behaviours that are unique to the application. Hence, these anomaly datasets are small and imbalanced, and a robust framework is required for such datasets. This paper proposes a Salient Textural Anomaly Proposal (STAP) framework to generate and classify salient textural proposals of regions of anomalies on metal surfaces. These anomalies have both salient and texture characteristics that are dependent on the properties of the metal surface. Furthermore, when observed across different lighting conditions, the anomalies in this AVI anomaly dataset have a small inter-class variance and large intraclass variance. The proposed STAP framework uses a Fourier transformation based technique to generate proposals of salient and textural anomaly regions. Transfer learning of Convolutional Neural Network (CNN) learned from a large dataset is used to train a linear Support Vector Machine (SVM) to classify the generated proposals. The proposed STAP framework performs the best when compared to state-of-the-art object recognition techniques on an AVI anomaly industrial dataset that has salient and texture anomalies on metal surfaces.