Abstract: Concerning the problem that handwriting feature extraction in the past taking the whole character as the research object can not detect stable features under lower relevance of the text, this paper puts forward a method of handwriting spurious dynamic feature extraction based on strokes, getting rid of the bondage of the nexus of contracts dependence. The method introduces the thought of probability statistics, using grid window to detect the spurious dynamic feature of the strokes, such as the tendency and the width variation. The method uses weighted Euclidean distance, weighted chi-square distance and weighted Manhattan distance to calculate the similarity of the handwriting identification. In the database involving HIT-MW and HIT-SW, the method achieves the top-1 identification accuracy of 95.9% and the top-10 accuracy of 99.5% under higher relevance of the text, and the top-1 identification accuracy of 91.9% and the top-10 accuracy of 99.0% under lower relevance of the text. Experiments show that the method of handwriting spurious dynamic feature extraction based on strokes can obtain a good effect under lower relevance of the text.

Key words: strokes, spurious dynamic feature, probability statistics, grid window, text relevance

摘要： 针对以往的以文字结体为研究对象的离线笔迹特征提取方法在文本相关度较低时无法获取稳定特征的问题，提出了一种以笔画为研究对象的笔迹伪动态特征提取方法，摆脱了结体依存性的束缚。引入概率统计思想，采用网格窗口提取笔画的运笔走势和宽度变化等伪动态特征。分别采用加权欧式距离、加权卡方距离和加权Manhattan距离计算笔迹相似度。在HIT-MW和HIT-SW库上进行实验，文本相关度较高时首选和前10选鉴别正确率分别为95.9%和99.5%；文本相关度较低时首选和前10选鉴别正确率分别为91.9%和99.0%。实验表明，以笔画为研究对象的笔迹伪动态特征提取方法在低文本相关度下仍能取得较好效果。

关键词: 笔画, 伪动态特征, 概率统计, 网格窗口, 文本相关度