基于曲线矩的手写体数字识别

RECOGNITION OF HANDWRITTEN NUMERALS BASED ON CURVE MOMENTS

不变矩理论自从提出以来,其应用及快速算法一直是一个比较热门的研究课题。以往的工作主要集中在利用面矩识别2D和3D物体。在本文中我们采用曲线矩的理论来识别手写体数字,这种方法有如下三个特点:1.与笔划的粗细无关;2.速度比面矩快得多;3.旋转、平移、比例不变。在本文中,我们使用Hu不变矩及Teague Zernike不变矩(TZMI)来识别手写体数字。同时改进了Teague Zernike不变矩,使其计算更简单,效果更好,更正了前人在使用Teague矩时的一个错误。还利用对各阶不变矩加权的办法,有效地减小了各阶不变矩的动态范围,大大地提高了识别效果。

After the theory of invariant moments has been proposed, its application and fast algorithm attract many researchers' attention. But in the past, the work concentrated on the recognition of 2D and 3D objects by using area moments. In this paper, we use the curve moments to recognize the handwritten numerals. This method has three characteristics: 1. independent to the stroke's width; 2. less computation than the method using area moments; 3. invariance under rotation, translation and scale. In this paper, we use Hu invariant moments and Teague invariant moments to recognize numerals. At the same time, we modify the formula of the Teague invariant moments and make it more simple and effective, correct one error of other's work. We also decrease the dynamic ranges among the invariant moments effectively by using the weight method.