针对目前纸币图像特征提取与分析方法缺少对相位结构信息有效描述的问题,提出一种基于四元共空间模式的特征提取算法.该算法首先采用四元矩阵描述纸币图像的相位与幅值信息,并对四元复合厄米特协方差矩阵进行对角化;然后将样本向量输入到复合四元滤波器中,并将分解得到的四元矩阵实部与虚部的方差作为纸币特征向量;最后采用神经网络作为分类器并在识别阶段引入拒识类.实验结果表明所提的算法不仅取得了较高的识别率,而且还能够满足纸币清分系统的实时性要求.%New feature extraction algorithm was proposed based on quaternion common spatial pattern in order to solve the lack of effective description of phase information in the banknote feature extraction and analysis. Firstly, the quaternion matrix was utilized to describe the phase information of the banknote image, and made diagonalization of quaternion composite Hermitian matrix. Secondly, the sample vector was input to the composite quaternion filter. The extracted feature vector was obtained by using the variance of the real part and imaginary part. Finally, the neural network was applied as classifier and the reject class was introduced in the banknote recognition. The experimental results illustrate that the proposed algorithm obtains high recognition rate and meets the real-time requirement of the banknote recognition system. The proposed algorithm has already been applied in a resource-constrained embedded system at the same time.
展开▼
