Adaptively Transfer Category-Classifier for Handwritten Chinese Character Recognition

Abstract

Handwritten character recognition (HCR) plays an important role in real-world applications, such as bank check recognition, automatic sorting of postal mail, the digitization of old documents, intelligence education and so on. Last decades have witnessed the vast amount of interest and research on handwritten character recognition, especially in the competition of HCR tasks on the specific data sets. However, the HCR task in real-world applications is much more complex than the one in HCR competition, since everyone has their own handwriting style, e.g., the HCR task on middle school students is much harder than the one on adults. Therefore, state-of-the-art methods proposed by the competitors may fail. Moreover, there is not enough labeled data to train a good model, since manually labelling data is usually tedious and expensive. So one question arises, is it possible to transfer the knowledge from related domain data to train a good recognition model for the target domain, e.g., from the handwritten character data of adults to the one of students? To this end, we propose a new neural network structure for handwritten Chinese character recognition (HCCR), in which we try to make full use of a large amount of labeled source domain data and a small number of target domain data to learn the model parameters. Furthermore, we make a transfer on the category-classifier level, and adaptively assign different weights to category-classifiers according to the usefulness of source domain data. Finally, experiments constructed from three data sets demonstrate the effectiveness of our model compared with several state-of-the-art baselines.