Support vector machine-based visualization for a multi-instance defect classification methodology in the dissimilarity space

Abstract

Automated visual inspection (AVI) systems for defect classification on the surface of objects involve sensing, preprocessing, representation, classification and post-processing. The first two stages produce either raw or preprocessed images, whereas the last one provides visualizations and actions based on the classification results. Regarding post-processing, this paper introduces a novel visualization technique—based on two statistically independent variables of a support vector machine (SVM), namely: the norm of the vector mapped by the SVM and the cosine of the angle between it and the polar vector of the separating hyperplane—which provides a hyperbolic visualization for an intuitive understanding of the defect classification results and an easy comparison against other classifiers. For the previous stages of the AVI system, we adopted a featureless methodology, relying on multi-instance learning (MIL) and dissimilarity spaces, to deal with inexactly labeled examples and eliminate the need for explicit feature extraction. The proposed SVM-based visualization and the featureless multi-instance methodology were evaluated with synthetic and real-world datasets, confirming the convenience of visually judging classification results, for an intuitive and interactive interpretation of the class label assignments, instead of only looking at small numerical differences in classification performances which, for very small orders of magnitude, may turn difficult to understand. In addition, high classification accuracies (above 0.97) were obtained under several grid sizes for MIL representation via block decomposition and when using different dissimilarity measures for all the considered datasets. The approach is scalable and generalizable across different datasets, making it a robust alternative for AVI tasks.