Chinese Character Recognition in Natural Scenes

this paper focuses in particular on the problem of Chinese characters recognition in natural scenes. Due to large variation in fonts, sizes, illumination, cluttered backgrounds, geometric distortions, etc., scene text recognition in the wild is a challenging problem. We proposed a novel method which based on Integral Channel Feature and pooling technology to extract informative features from scenes images. We concatenated many different low-level features and selected typically features to represent the Chinese characters. In this work, we make use of Support Vector Machines as the classifier, and rank the features by the weights of training model in LinearSVM. Thus features representation of characters is compact and it is effective to express distinctive spatial structures of text character. At the same time, for comparative purpose, we evaluated approach extensively on two standard dataset (ICHAR03, Char74K). Because of the absence of Chinese character datasets, we took 311 photos by cellphone and digit camera in Guangzhou and cropped them to pieces manually to form two Chinese datasets. Our experiment results denote that our proposed technology was performed better than the current state-of-the-art methods.