Using metamorphic relations to verify and enhance Artcode classification

Software testing is often hindered where it impossible or impractical to determine the correctness of behaviour output software under test (SUT), a situation known as oracle problem. An example an area facing problem automatic image classification, using machine learning classify input one set predefined classes. approach that alleviates metamorphic (MT). While traditional examines individual cases, MT instead relations amongst multiple executions cases and their outputs. These are called (MRs): if MR found be violated, then fault must exist in SUT. This paper classifying images containing visually hidden markers Artcodes, applies verify enhance trained classifiers. further two MRs, Separation Occlusion, reports on capability verifying classification one-way analysis variance (ANOVA) conjunction with three other statistical methods: t-test (for unequal variances), Kruskal–Wallis test, Dunnett’s test. In addition our previously-studied classifier, used Random Forests, we introduce new classifier uses support vector machine, present its MR-augmented version. Experimental evaluations across number performance metrics show augmented classifiers can achieve better than non-augmented also analyses how enhanced obtained.