改进的HOG和SVM的硬笔汉字分类算法

Improved HOG and SVM hard-pen Chinese character classification algorithm

针对目前HOG提取汉字特征时存在维度过大、特征边缘化的问题,结合汉字网格技术提出一种基于网格的分层HOG特征提取算法。以特征块无重叠的方式提取一层HOG特征,提取底层均匀块的梯度特征,融合两层特征。该算法可有效提取汉字轮廓特征,降低特征维度。在此基础上,提出较为完善的中小学硬笔汉字分类评价框架流程,结合线性PCA降维,采用SVM分类器,实现硬笔汉字的三级分类。通过多个汉字结构的分类实验,验证了该算法的准确性和有效性。

To solve the problem of large dimensions and marginal features in the current HOG extraction of Chinese character features,a grid-based hierarchical HOG feature was proposed in combination with the Chinese character grid technology.A layer of HOG feature was extracted in a non-overlapping way.The gradient features of the underlying homogeneous block were extracted.Two layers of features were fused.The algorithm was used to extract contour features of Chinese characters and reduce feature dimensions.On this basis,a relatively perfect evaluation framework of hard-pen Chinese character classification was put forward.Among them,linear PCA dimension reduction was combined,and SVM classifier was used,to achieve the three-level classification of hard-pen Chinese characters.The accuracy and effectiveness of the algorithm were verified by the classification experiments of several Chinese characters.