CBIR Based Testing Oracles: An Experimental Evaluation of Similarity Functions

Abstract

Content-Based Image Retrieval (CBIR) systems constitute an innovative approach to store, to compare and to query images in a database. Visual aspects such as color, texture or shape are used to perform such operations. Recently, CBIR concepts were applied to build testing oracles for image processing programs, where test verdicts (approval/disapproval) are based on similarity measures between images produced by the program and reference images. However, the results of a CBIR system may vary depending on the components employed in the system (feature extractors and similarity functions), and few studies assessing this influence have been found in the literature. Our aim is to present an empirical analysis of ten similarity functions in CBIR systems within the context of software testing with graphic outputs. A case study with images obtained from a computer-aided diagnosis system in mammography indicated some variability among image test verdicts (approval/disapproval) according to the similarity function choice. The case study also indicates the existence of some clusters of similarity functions with high correlation coefficients.