高光谱图像结合机器学习方法无损检测猕猴桃

Nondestructive detection for kiwifruit based on the hyperspectral technology and machine learning

为实现对过量使用1-MCP化学保鲜剂猕猴桃快速、无损检测,提出高光谱技术结合机器学习建立识别模型的检测方法。首先对空白猕猴桃和过量化学保鲜猕猴桃在865.11～1 711.71 nm范围内进行高光谱数据采集。然后选用标准正态变量变换方法预处理原始光谱数据以去除噪声,采用波段比算法增强图像,数学形态学算法提取感兴趣区域,进而计算光谱平均值。最后采用主成份分析(PCA)、竞争性自适应加权(CARS)方法对全光谱数据(FS)进行特征提取,去除干扰项;以PCA和CARS提取的特征量和FS数据作为输入,结合偏最小二乘(PLS)和支持向量机(SVM)建立12个识别模型。试验结果表明,基于PLS和SVM建立的识别模型均能够有效检测过量化学保鲜猕猴桃,其中CARS-SVM模型性能最好,平均正确识别率达100%,运行速度最快,仅为0.015 348 s,满足工程实践中实时性高的要求,为快速、无损检测猕猴桃果品安全提供理论支撑。

In order to detect on the kiwifruit with excessive 1-MCP quickly and nondestructively,this paper proposes the detection method based on the hyperspectral technology and machine learning.Firstly,representative samples are got by experiment,and hyperspectral images of sample are obtained from 865.11 nm to 1 711.71 nm.Secondly,the original hyperspectral data are pre-processed by applying the standard normal variable transformation method to denoise samples,then the region of interest are extracted by using the mathematical morphology algorithm after image enhancement using band ratio algorithm to calculate the average spectral value of each sample.Finally,the principal component analysis(PCA)and competitive adaptive reweighted sampling(CARS)algorithms are adopted to extract effective features from the full spectral(FS)data and remove interferences,and 12 recognition models based on partial least squares(PLS)and support vector machine(SVM)are established using the PCA,CARS and FS data respectively.Experimental results showed that both PLS and SVM model were able to recognize the 1-MCP kiwifruit and normal kiwifruit,and the performance of the CARS-SVM model was the best among models,and the average recognition rate and elapsed time of which was 100%and 0.015 348 s,which meet the high real-time demand well.