Joint Spatial and Radical Analysis Network For Distorted Chinese Character Recognition

Recently, a novel radical analysis network (RAN) has been proposed for Chinese character recognition (CCR). The key idea is treating a Chinese character as a composition of radicals rather than a single character class. Compared with traditional learning ways, two serious issues in CCR, i.e., enormous categories and limited training data, can be effectively alleviated. In this paper, we further excavate the potential capability of RAN. First, we validate RAN can reduce the equivariant requirement of regular convolutional neural network (CNN) owing to finer modeling and a local-to-global recognition process, especially considering the rotation transformation. This modeling approach of RAN can be regarded as one instance of compositional models. Second, we propose a joint spatial and radical analysis network (JSRAN) to handle more general situation in which the test data includes kinds of affine transformations. No matter for rotated printed Chinese character or natural scene, JSRAN can outperform RAN and traditional CNN. Finally, according to visualization analysis, we empirically explain why JSRAN can yield a remarkable improvement.