基于高光谱图像的红豆品种GA-PNN神经网络鉴别

Identification of Red Bean Variety with Probabilistic GA-PNN Based on Hyperspectral Imaging

提出一种基于高光谱图像技术的红豆品种鉴别方法。利用高光谱成像系统采集江苏、安徽、山东的3个品种共162个红豆样本高光谱图像数据,通过ENVI软件提取出红豆中感兴趣区域的平均光谱作为该样本原始光谱信息,利用SG多项式平滑对原始光谱数据进行去噪处理,由于高光谱数据信息量大,冗余性强,故需对高光谱数据进行降维,采用了连续投影算法进行特征波长选择,根据交叉验证均方根误差确定最佳特征光谱的个数为9,采用主成分分析法和独立分量分析算法进行特征波长提取,经过PCA处理,根据方差累计贡献率大于85%的标准选出7个特征波长,ICA分别提取了7、10、17个特征波长,通过测试集验证,选出17个最佳特征波长。最后分别将优选出的特征波长和提取出的最优主成分作为模型的输入。建立概率神经网络(PNN)模型测试后发现结果没有达到预期精度,引入遗传算法(GA)优化的PNN神经网络的阈值,并对隐含层节点进行最优选择。通过测试试验,所有的模型识别正确率均高于85%,其中SPA-GA-PNN模型的效果最佳,识别正确率达到了97.5%。

A method to identify different varieties of red bean based on hyperspectral imaging technology was proposed. The hyperspectral imaging system with spectrum range of 390 ~ 1 050 nm was used to capture the hyperspectral images of 162 red bean samples,which were collected from three different areas（ Anhui,Shandong and Jiangsu Provinces）. ENVI software was adopted to determine the region of interest（ ROI） in the hyperspectral image and extract the hyperspectral data by averaging the reflectance from all the pixels in the ROI images,and the original spectra were preprocessed by Savitzky-Golay（ SG）smoothing. As there was a large number of noise and redundant information in the raw hyperspectral images and hyperspectral data,some data processing methods should be used to remove the noise,accelerate the processing efficiency and improve the performance of the models. The method of feature extraction was SPA,the number of characteristic wavelengths was determined as 9 by using the leave-oneout cross-validation. The methods of feature selection were PCA and ICA. According to the standard of the cumulative contribution rate of variance was more than 85%,seven characteristic wavelengths were selected. Through test and verification,17 was the best number of characteristic wavelengths of ICA.Finally,the selected optimal characteristic wavelengths and principal components were used as the inputs of the model. However,the results did not meet the expected accuracy,the threshold of PNN neural network and hidden layer nodes were optimized by GA. The recognition rate of the model was higher than85%,and the recognition rate of the highest SPA-GA-PNN model reached 97. 5%. The results demonstrated that it was feasible to use hyperspectral imaging technology for the identification of red bean variety. PNN neural network model can identify red bean variety fast,effectively and nondestructively and provide theoretical basis and technical means for the realization of red bean variety identification based on hyperspectral image technology.