Object classification using Radon transform and Convolutional Neural Networks

This paper presents a novel approach to perform object classification based on combining the conic Radon transform (CRT) and the convolutional neural networks (CNN). The CRT generalizes the classical Radon transform (RT) to detect conic sections in images. We have build, in this work, a descriptor combining the extracted features by the circular, parabolic and linear RT. The obtained descriptor is then entered as an input into the convolutional layers. In order to evaluate the efficiency of this approach, we have carried out experiments on the ETH80 data set. The obtained results show the performance of this new approach which joins together extraction of features of different (circular, parabolic and linear) shapes and the convolutional neural networks. Since the use of large database is required for deep learning techniques to get accurate results, we investigate in this work the use of small database on simple CNN models. Our technique achieved competitive classification accuracy and superior computational complexity compared to the state-of-the-art methods.