Non-Negative Matrix Factorization Based Off-line Handwriting Identification

Handwriting identification is one of the popular research subjects in the area of computer vision and pattern recognition. It aims to judge the identity of a writer by comparing the same character written by different people and analyzing the writing style. Most of the traditional handwriting identification methods tried to extract the handwriting texture features within the whole paragraph. In order to improve the identification accuracy rate, this paper applies non-negative matrix factorization to the off-line handwriting identification of individual characters.Prior to NMF process, two non-negative matrixes need to be initiated at random. In order to increase the stability of the algorithm, this paper uses the initialized values from high dimensional space onto the low dimensional space. Then NMF is applied to the representative characters, which are selected out of a section of characters, to get the handwriting picture's sub-eigenspaces and eigenvectors for each charactor. Mapping the test samples onto the sub-eigenspaces, get eigenvectors, and calculate the angle relativity and k-nearest neighbor between the eigenvectors and the values of test samples, then classify the handwriting picture. The identification accuracy rate for the complicated Chinese characters that have left-right structure is relatively higher. This results show that NMF outperforms the traditional PCA-base representation. Therefore, non-negative matrix factorization technology has potential in the analysis of writer identification.