Multi-scene ancient Chinese text recognition with deep coupled alignments

Abstract

The task of multi-scene ancient Chinese text recognition (MACR) is challenging due to large-scale categories, high intra-class variance and inter-class similarity and complicated backgrounds. Little effort has been devoted to MACR research due to insufficient datasets and language barrier. Because the sub-dataset generation process of sub-dataset is mutually blind, there are discrepancies in the class category number, deep feature representation and class center distribution after the dataset statistics and character analysis are performed. The general deep learning method that assumes that data are independent and identically distributed is inappropriate. The deep coupled alignments (CA) module based on domain adaptation is presented to alleviate domain and class center shifts. In addition, a cross-domain fusion (CF) module is proposed to mitigate negative transfer in partial domain adaptation by updating the target domain with the full-class and augmenting the source domain with pseudo labeled samples. Extensive experiments of the proposed method are conducted, and the results illustrate the superiority of CA–CF to previous methods in terms of the model size and recognition accuracy.