基于SVM的多特征唐卡头饰分类

Classification of Thangka headdress based on multi-feature and SVM

为了能够适应寺院、图书馆、文化馆、唐卡数字化保护研究机构等用户精确检索的需要,对唐卡头饰进行分类,从而实现语义标注、语义检索,提高唐卡检索的精准度。针对已有的唐卡头饰分类方法具有分割困难或交互繁琐等缺点,不利于实际应用,因此本文提出了基于SVM的多特征唐卡头饰分类方法。首先利用小波分割方法、FFT分割方法对头饰进行分割,获取头饰轮廓信息,分别提取Hu不变矩、傅里叶不变矩、Zernike矩和频率谱形状特征;其次合并这两类形状特征,得到合并形状特征,与头饰的颜色特征联合成多特征;最后SVM训练后分类。与其它方法相比,本文提出的分类方法具有分割效果好且交互简单等优点,可以达到对头饰分类实际应用的要求。

In order to meet the precise retrieval needs of temples, libraries, cultural centers, and research institutions for the digitized protection of Thangka, the classification of headdress can be used in semantic annotation,semantic retrieval to improve the precision of Thangka retrieval. The past methods in classifying headdress, which is unfavorable for practical application,have some shortcomings such as segmentation difficulties or interaction trivia, therefore, the paper proposed the classification algorithm based on SVM and multiple features. Firstly, obtain the headdress contour using wavelet and FFT segmentation algorithm, and extract the Hu invariant moment, Fourier invariant moment and Zernike moment and frequency spectrum shape characteristics respectively; secondly, get shape feature by combining the two kinds of contour features,and it unites color features to generate the multiple features; Finally, train SVM and test classification. Compared with other classification methods, the proposed method by the paper has good segmentation effect and simple interactions, which can meet the requirements of headdress classification in the practical application.