Study On Dimension Reduction Characteristics Of Online Handwritten Sanskrit Tibetan Distinction Based On Similarity

Feature dimension reduction plays an important role in the field of pattern recognition. The process of online handwritten Tibetan fine handwriting recognition includes preprocessing, feature extraction, feature dimension reduction and recognition. Sanskrit Tibetan has many similar words. According to the regularity of low dimensional space data to analysis and find similar words between different similar for increasing recognition rate.Firstly, to reduce initially the dimension of the character feature matrix using singular value decomposition for initial feature vector in low dimension space. To construct similarity matrix space by analyzing the similarity between characters feature matrix in the low dimensional space. On the basis of the similarity matrix space, the initial feature vectors of the low dimension space are classified or analyzed by using the spectral clustering. Then, to analyses the classification information of similar word categories and no similar word categories. Based on the category relation between the similar words, the classification character matrix is analyzed. Based on the category relation between the similar words, the classification character matrix is analyzed. With PC A and PCA+LDA methods to reduce finally the dimension of the character feature matrix, and then improve the convolution process for the LDA to ensure that there is a good ability to distinguish between similar words and get the class information of the descending of the reform matrix.Finally get the results of the character category and a variety of forms of the reduction of the matrix results, these reduced the restoration of the matrix to distinguish between similar words. The experiment makes six forms of dimension reduction for the original character matrix. (1) With PCA to reduce the dimension of the character feature matrix; (2) With PCA+LDA to reduce the dimension of the character feature matrix; (3) SVD for preliminary dimension reduction, and then PCA for dimension reduction; (4) SVD for preliminary dimension reduction, and then PCA+ LDA for dimension reduction; (5) With SVD preliminary dimension reduction to obtain the preliminary dimension reduction matrix, next, spectral clustering for dimension reduction, and then PCA+LDA for dimension reduction; (6) after the initial dimension reduction matrix obtained by SVD, the spectral clustering classification is carried out, and the original data classification matrix is reduced by PCA+ LDA. Finally, the reduced dimension matrixes of six kinds of results are obtained. To classify the results of the six types of dimension reduction matrix to get the new classification results. Comparison of the two classifications, the first classification of similar characters in the same category, the new classification results in similar words are separate from the situation.classification process.Compared with the six cases with the degree of similarity words in the two classification process is separated. The experimental results show that the sixth methods can distinguish the complex similar words.