Recognition of OBIC's Variants by Using Deep Neural Networks and Spectral Clustering

Oracle bone inscriptions (OBIs) are the origin of Chinese characters and play a pivotal role in the study of Chinese civilization and the world civilization. The automatic recognition of OBI character (OBIC) images is very import to the research and promotion of OBI culture. However, a large amount of these ancient characters have variants with totally different appearance, which brings very serious negative impact on the OBI studies. In this paper, we proposed to recognize variants of OBICs by combining deep convolutional neural networks (DCNNs) with spectral clustering (SC). The former is employed to provide accurate descriptions for OBIC images, and the latter is used to find variants of each OBIC class. More specifically, the pretrained ResNet50 is exploited to obtain image features, and the normalized graph cuts is employed to find variants. Besides, a label propagation algorithm is used to find the label of test OBICs based on the clustering results. The proposed method is tested on an OBIC image set, in which all images are cropped from OBI rubbing images. Experimental results have shown that our method has the ability to recognize OBIC's variants.