A multiscale shape descriptor based on diferential entropy

We propose a shape descriptor based on the diferential entropy of the multiscale curvature (MEC) that improves the discrimination ability of shape representation. We conduct multiple-class classification experiments on images from three public datasets: Kimia99 (99 shapes / 11 classes), Flavia leaves (1907 shapes / 32 classes) and MPEG7-CE (1400 shapes/ 70 classes). The quantitative evaluation of the classification results in terms of Precision, Recall and F1- measures show that MEC improved shape description for the MPEG7 and Flavia datasets. The qualitative analysis of MEC performed by the pairwise euclidian distance matrices confirmed that it is competitive for shape description when compared to the normalized multiscale bending energy descriptor.