经验模型分解在手写体数字识别中的应用

The Application of Empirical Mode Decomposition on Handwritten Numerals Recognition

将经验模型分解方法应用于手写体数字识别,提出了一种新的识别方法.该方法基于字符的轮廓信息特征,具有平移不变性、缩放不变性、旋转不变性.本文基于分段线性逼近方法提取手写体数字图像的外轮廓,由外轮廓数据构造一个信号s(t),对s(t)进行经验模型分解以抽取它的第一个内蕴模式函数并计算该内蕴模式函数的瞬时频率,选取5个较大的瞬时频率值以及它们之间的时间间隔作为9个特征值;然后对s(t)的波形再重采样16个点作为16个描述s(t)的波形的特征值;最后,采用基于反向传播算法的有多个多输入单输出形式的三层前馈神经网络进行分类与识别.实验结果表明该方法获得了97.5%以上的正确识别率.

This paper presents a novel method for handwritten numeral recognition based on empirical mode decomposition. It is invariant under translation, rotation, and scaling transformation. Based on contour tracing by piecewise linear approximations method, the closed contour of the numeral in an image is extracted, a feature signal, s（t）, is constructed out of the contour data. By decomposing s（t） with the empirical mode decomposition, the first intrinsic mode function （imf） and its instantaneous frequencies is computed. By detecting the five largest instantaneous frequencies and calculating the corresponding distances between each neighboring pairs of them （there are four distances）, we get a feature vector of dimension 9. To descript the signal＇s wave form, another feature vector of dimension 16 is obtained by re-sampling the signal s（t） to 16 points uniformly, therefore a feature vector of dimension 25 is formed by combine these two feature vectors. Finally, based on backpropagation algorithm, multi- multiple inputs one output type three layers forward neural network is devised to classify the handwritten numerals of { 0, 1,…, 9} and 1000 samples from CENPARMI database are employed to train this classifier. The experimental results show that the method provide an encouraging match rate of 97%.