一种用于玉米品种鉴别的近红外光谱频率选择方法

提出了一种近红外光谱的频率选择方法用于玉米品种鉴别。首先确定一种衡量特征鉴别能力的准则函数,然后根据该准则函数逐步选出适合于分类的特征频率,并通过去除各频率特征之间的相关性使得优选出的频率特征包含尽可能多的品种类间差异信息,优先选择方差较大的频率特征以减弱噪声的影响。实验结果表明,频率选择大幅度地改善了识别效果,仅使用30维频率特征即可达到94.16%的识别率。随机模拟实验显示,优选出的频率特征的识别效果对频率的小幅随机扰动不敏感,验证了本方法的鲁棒性。

A frequency selection method of NIR spectroscopy for discrimination of maize seed varieties

A frequency selection method of NIR spectroscopy was proposed in the present paper for discrimination of maize seed varieties. A criterion function was defined to evaluate the discriminative ability of NIR spectroscopy at different frequencies, and then features of maize seed varieties were extracted accordingly for further processing. By eliminating correlation between features at different frequencies, the selected features are guaranteed to contain as much information of inter-variety difference as possible. Also, features with larger variances are preferred to suppress the impact of noise. Experiment results demonstrate that our frequency selection method can achieve high recognition rate with less spectroscopy features than traditional methods. Specifically, a recognition rate as high as 94.16% can be attained with NIR spectroscopy with only 30 frequencies. Simulation results show that recognition rate of NIR spectroscopy at selected frequencies is stable with small disturbance of frequencies, which verifies the robustness of the authors' method.