Particular Writer’s Keyword Spotting Using Geometric Information And SIFT Features

There is a practical demand of particular writer’s keyword spotting in Chinese document images. However, the relative research is far from enough. Especially, large variety of Chinese characters and arbitrary of handwriting style leads to a big challenge for keyword spotting in Chinese handwritten documents. A new method based on combination of character geometric information and SIFT features, is proposed for detecting keywords of particular writers.SIFT has been proven a stable and distinctive local feature, which can perform well in distinguishing different handwriting. The SIFT features respectively extracted by document images and keyword images can be simply matched by Kd-tree algorithm. In order to avoid the difficulty of character segmentation, a method of sliding window is applied to improve the efficiency of feature matching.Combined with character geometric information, a geometry constraint graph can be established. By the use of maximum clique matching, the proposed method can find the maximum clique in the geometry constraint graph and effectively remove miss-matching feature points and therefore improve the precision rate of detection. The keyword spotting result can be obtain after filtrating the windows by judging the number and centroid deviation of the matched feature points in them. Finally, the precision of the detecting result can be further improved after post-processing based on the similarity of images.A big amount of experiments are taken on the HWBD2.1from CASIA. The results in handwriting document images show that the method can efficiently detect keywords of particular writers and remain high recall rate and high precision rate.