Banknote recognition as a CBIR problem

Automatic banknote recognition is an important aid for visually impaired users, which may provide a complementary evidence to tactile perception. In this paper we propose a framework for banknote recognition based on a traditional Content-Based Image Retrieval pipeline: given a test image, we first extract SURF features, then adopt a Bag of Features representation, finally we associate the image with the banknote amount which ranked best according to a similarity measure of choice. Compared with previous works in the literature, our method is simple, computationally efficient, and does not require a banknote detection stage. In order to validate effectiveness and robustness of the proposed approach, we have collected several datasets of Euro banknotes on a variety of conditions including partial occlusion, cluttered background, and also rotation, viewpoint, and illumination changes. We report a comparative analysis on different image descriptors and similarity measures and show that the proposed scheme achieves high recognition rates also on rather challenging circumstances. In particular, Bag of Features associated with L2 distance appears to be the best combination for the problem at hand, and performances do not degrade if a dimensionality reduction step is applied.