A Novel Approach of Deep Convolutional Neural Networks for Sketch Recognition

Abstract

Deep Neural Networks (DNNs) have recently achieved impressive performance for many recognition tasks across different disciplines including image recognition task. However, most of existing works on deep learning for image recognition focus on natural image data (photo-based images) and not on sketches. Moreover, most of existing works on sketch classification are based on hand crafted feature representations. In this paper, we propose to train a convolutional neural network for sketch recognition using the TU-Berlin sketch dataset composed of 250 object categories with 80 images each. We find that training a CNN with a proper data-augmentation and a multi-scale multi-angle voting technique can achieve an accuracy of 75.43%, which surpasses human-level performance in the standard sketch classification benchmark and significantly outperforms state-of-the-art sketch recognition methods.