多尺度特征和神经网络相融合的手写体数字识别

Handwritten Numeral Recognition Based on Multi-scale Features and Neural Network

针对传统特征提取方法无法有效解决书写随意性的干扰问题,提出了一种多尺度特征和神经网络相融合的手写体数字识别方法。首先提取手写体数字二值图像的轮廓、笔画次序等结构特征,并旋转坐标轴,提取多角度结构特征;然后将字符从中心点到外边框划分为K层矩形子层,提取每层图像的灰度特征,最后以两种多尺度特征构建神经网络模型,并预测测试集合样本。将该算法实际用于以MNIST字体库构建的两个数据集识别,其精度高达99.8%,并能有效降低倾斜等手写字体的随意性影响。

Aiming at the problem that tradition handwritten numeral recognition method can not solve the interference from writing arbitrary, a new handwritten numeral recognition method was proposed based on nmulti-seale features and neural network. Firstly, two structural features of outline and strokes were extracted, and multi-angle structural features were extracted by rotating the datum line. Second, Multi-level grayscale pixel features were extracted by dividing the ima- ge to K sub-layer from the inside out. Thirdly, BP neural network model was build based on the two features. Lastly, new method was used for The MNIST font library, and the prediction precision reached 99. 8%. The result shows that new algorithm can effectively reduce the impact of tilt.