Recognition Of Ancient Chinese Characters Based On Hybrid Kernel LS-SVM

Ancient Chinese characters recorded a large amount of data of polity, economy, history and so on, which have a very high historical value. However, the recognition of ancient Chinese characters is very difficult. On the one hand, the strokes are irregular and the number of variants is large; on the other hand, ancient Chinese characters are always appeared in the forms of inscription and handwriting, which lead to serious damage. Therefore, it is conducive not only to the inheritance and development of the national culture, but also to the circulation and collection of electronic process of ancient books by utilizing image processing technology to recognize ancient Chinese characters.Since a large number of variants and local deformations exist in ancient Chinese characters, the existing methods are difficult to obtain accurate results. Support vector machine (SVM) with the abilities of strong generalization and anti-noise has been widely applied in image recognition to deal with small-sample problems. In this paper, the hybrid kernel least squares support vector machine (LS-SVM) is combined with feature extraction or Curvelet transform to realize the ancient Chinese characters recognition. The main work and conclusions are as follows:1. Due to the high similarity among the ancient Chinese characters, the misclassification is inevitable. In this paper, hybrid kernel weighted LS-SVM is used to conduct classification instead of the traditional SVM, which can reduce the negative influence of abnormal samples and improve the classification accuracy.2. A Multi-feature fusion method of time domain is proposed in the paper. The component structure feature and the overall density feature are extracted as the global feature, which has the advantages of strong robustness and low complexity of algorithm. The local point density characteristics in pseudo 2D elastic grid and the grid stroke feature are extracted as the local features, which can ensure the good absorption ability for local deformation. The extracted features are sent into the classifier after fusing operation.3. Since the strokes of ancient Chinese characters are irregular curves, the classification rate is low. According to this problem, the fast discrete Curvelet transform of the second generation is used to extract the frequency feathers of ancient Chinese characters. In addition, a multi-feature fusion method in frequency domain is proposed in the paper. Firstly, the multi-resolution decomposition of ancient Chinese characters is conducted by using the fast discrete Curvelet transform of the second generation. Secondly, the gray level co-occurrence matrix of each subimage in different resolutions is calculated respectively. Thirdly, the multi-feature fusion operation with the feature parameters obtained from each subimage is conducted to fuse a high-dimension vector serial vector for the PCA dimension reduction. Then, the low-dimensional vectors are sent into the classifier. Experiment results show the effectiveness of the proposed method.