ToonNet: a cartoon image dataset and a DNN-based semantic classification system

Abstract

Cartoon-style pictures can be seen almost everywhere in our daily life. Numerous applications try to deal with cartoon pictures, a dataset of cartoon pictures will be valuable for these applications. In this paper, we first present ToonNet: a cartoon-style image recognition dataset. We construct our benchmark set by 4000 images in 12 different classes collected from the Internet with little manual filtration. We extend the basal dataset to 10000 images by adopting several methods, including snapshots of rendered 3D models with a cartoon shader, a 2D-3D-2D converting procedure using a cartoon-modeling method and a hand-drawing stylization filter. Then, we describe how to build an effective neural network for image semantic classification based on ToonNet. We present three techniques for building the Deep Neural Network (DNN), namely, IUS: Inputs Unified Stylization, stylizing the inputs to reduce the complexity of hand-drawn cartoon images; FIN: Feature Inserted Network, inserting intuitionistic and valuable global features into the network; NPN: Network Plus Network, using multiple single networks as a new mixed network. We show the efficacy and generality of our network strategies in our experiments. By utilizing these techniques, the classification accuracy can reach 78% (top-1) and 93%(top-3), which has an improvement of about 5% (top-1) compared with classical DNNs.