Image descriptors analysis supported by Information Visualization with application in automatic classification

Abstract

Automatic image classification strongly depends on image representation, commonly made using visual descriptors. Several works compare the variety of the available ones, aiming to guide analysts in which one to choose for each scenario, but they consider only numerical accuracy results, which may limit the understanding of reasons for these performance differences. This paper employs Neighbor Joining similarity trees to visually analyze three image descriptors, focusing on their application in an automatic classification scenario. The results demonstrated that these trees provide means to comprehend important information about the descriptors, such as how they describe images characteristics, which image aspects they focus in the representation, and which criteria they consider to distinguish images into different classes, revealing their strengths and limitations regarding the representation of a specific categorization scheme. We believe such analysis may help specialists to better choose which descriptor is adequate to be used in this data mining task.