基于高光谱的茄子外部缺陷检测

利用高光谱(900～1700nm)对完好、木栓化和烂果茄子进行识别研究。共采摘了252个茄子样本,包含完好茄子170个,木栓化茄子60个和烂果茄子22个,利用高光谱成像系统采集完好、木栓化和烂果3种区域一共252个样本的高光谱图像,然后提取合理的感兴趣区域(ROI)获得样本光谱数据。采用多种预处理方法进行光谱预处理,建立偏最小二乘(partial least squares method,PLS)判别分析模型,结果表明,经normalize预处理后模型的预测效果最好,因此选择normalize作为预处理方法。基于预处理后的光谱数据,采用连续投影法(SPA)、回归系数法(RC)和竞争性自适应重加权算法(CARS)提取特征波长,并分别建立偏最小二乘(PLS)和多元线性回归(MLR)判别模型进行研究。结果表明:CARS-MLR模型对3种类型样本鉴别效果最佳,其校正集决定系数Rc2为0.94,预测集决定系数Rp2为0.90,RMSEC和RMSEP分别为0.19和0.21,预测集判别准确率达到96.82%。本研究采用高光谱可以对完好、木栓化和烂果茄子进行有效鉴别,为茄子的缺陷无损检测提供了理论参考。

Detection of Eggplant External Defects Using Hyperspectral Technology

Hyperspectral technology (in the range of 900~1700 nm) was employed to distinguish sound, suberized and rotten eggplants. A total of 252 eggplant samples were collected, including 170 sound eggplants, 60 suberized eggplants and 22 rotten eggplants. The hyperspectral imaging system was used to acquire hyperspectral images of 252 eggplant samples in the three types of areas (sound, suberized and rotten fruit regions), and a reasonable region of interest (ROI) was extracted to obtain the spectral data. A variety of preprocessing methods were used for spectral pretreatment and a partial least squares (PLS) discriminant analysis model was established. The results show that the prediction was the best after the model was subjected to normalization pretreatment. Therefore, normalization was selected as the pretreatment method, based on the pre-treated spectral data, the characteristic wavelengths were extracted by successive projections algorithm (SPA), regression coefficient (RC) and competitive adaptive reweighted sampling (CRAS) methods, and partial least squares (PLS) and multiple linear regression (MLR) discriminant models were established for analysis. The results showed that the CRAS-MLR model discriminate most effectively the 3 types of samples, with the coefficient of determination for the calibration set (Rc2) as 0.944, the coefficient of determination for prediction set (Rp2) as 0.901, and the RMSEC and RMSEP as 0.199 and 0.213, respectively. The discriminant accuracy of the prediction set was 96.8%. In this study, hyperspectral technology can be used to distinguish effectively the sound, suberized and rotten eggplants, which provides a theoretical reference for the nondestructive detection of eggplant defects.