Worm-like image descriptor for signboard classification

Abstract

© 2020 for this paper by its authors. We introduce a special image descriptor that is well suited for classification of images containing various inscriptions. In order to demonstrate effectiveness of the proposed solution we provide evaluation of a system based on the introduced descriptor on commercial building facade photographs grouping problem according to the type of services provided. Our system achieved state of the art performance (0.28 in averaged F1) over classical CNN-based methods and a composite baseline.