Sketch-a-Net that Beats Humans

Deep Neural Networks (DNNs) have recently outperformed traditional object recognition algorithms on multiple large-scale datasets, such as ImageNet. However, the model trained on ImageNet fails on recognising the sketches, because the data source is dominated by photos and all kinds of sketches are roughly labelled as 'cartoon' rather than their own categorises (e.g., 'cat'). Most of sketch recognition methods still heavily rely on the hand-craft feature extraction techniques, thus it is interesting to know whether DNNs can eliminate the needs of specific feature engineering in this area, and how the architecture is designed for sketch recognition purpose. To the best of our knowledge, it is the first deep neural network model for sketch classification, and it has outperformed state-of-the-art results in the TU-Berlin sketch benchmark. Based on that, we outline a sketch image retrieval system in a unified neural network framework.