Overlaid Chinese Character Recognition via a Compact CNN

Abstract

Great successes have been enjoyed in the previous work for Chinese character recognition (CCR), however, few impressive works have been done about the recognition of Chinese characters with complex backgrounds. This paper focuses on the recognition of overlaid Chinese characters - the Chinese characters embedded in images or videos - which are often with complex backgrounds and of diverse typefaces and styles. In this paper, we present a high-performance recognizer based on the deep convolutional neural network (CNN). To train the CNN, a large number of character images are first collected by the synthetic way. By fully considering the input size, depth, width, and filter sizes of a network, we present multiple candidate models with compact network architectures. Comprehensive comparison experiments are carried out to help us select the model, which requires only 13.6M for storage (3.6M parameters) and takes only 0.038 s for recognizing 3755 character images on a GPU. The experimental results shows that the model achieves the recognition rate of 99.77% on the test set, and a good generalization performance is also validated on the dataset of typefaces not included in the training set. Besides, the extensive comparison experiments presented in this paper might give lights into the formation of deep CNN.