以贮藏1d的碰撞损伤猕猴桃、挤压损伤猕猴桃和无损猕猴桃为对象,分别建立了猕猴桃的Fisher判别模型、BP神经网络判别模型与最小二乘支持向量机(LSSVM)判别模型,综合比较了采用全光谱波长(FS)、主成分分析(PCA)提取特征变量与连续投影算法(SPA)优选特征波长作为各模型输入变量时,对各模型判别效果的影响.研究结果表明,SPA优选特征波长相比于PCA和FS有较明显的优势;3种判别模型均能基本满足实际要求,且LSSVM模型的识别性能最佳,其中SPA-LSSVM模型对预测集碰撞损伤样品、挤压损伤样品与无损样品的正确识别率分别达到100％、95％和100％,总的正确识别率为98.2％.%To detect bruised kiwifruits from intact kiwifruit early and effectively, near infrared diffused reflectance spectroscopy technology combined with Fisher discriminant function, BP neural network and least square support vector machine ( LSSVM) were applied to discriminate collided kiwifruit, pressed kiwifruit and intact kiwifruit after storage of 1d, respectively. Effectiveness of the discriminant model using full spectrum ( FS ) , feature variables based on principal component analysis ( PCA ) and characteristic wavelength by successive projection algorithm ( SPA ) was compared and evaluated. The results showed that SPA gave the best advantage compared with methods of FS and PCA. Three models all had an acceptable accuracy, especially LSSVM model had the optimal recognition performance. SPA - LSSVM had an accuracy rate of 100% , 95% and 100% for identifying collided samples, pressed fruits and intact ones respectively, and the discriminant accuracy rate for total samples was 98. 2% .
展开▼
