应用高光谱技术及MLSPTSVM模型检测热损伤大豆

进口大豆在运输过程中极易因贮藏温度过高而造成热损伤,加剧大豆蛋白及油脂的品质恶化,对大豆质量造成影响。本文利用高光谱图像技术和多元最小二乘递归投影孪生支持向量机(multiple least squares recursive projection twin support vector machine,MLSPTSVM)对大豆的热损伤进行检测。应用高光谱图像采集系统,在400~1000 nm范围内获取正常大豆、轻度热损伤、重度热损伤大豆的光谱图像,采用多种预处理方法进行光谱预处理,对预处理方法提高模型检测性能的有效性进行分析。结果表明,多元散射校正预处理搭配线性核的MLSPTSVM模型,原始光谱数据搭配非线性核的MLSPTSVM模型,均能达到100%检测准确率,相较于经典检测模型具有显著优势。在实验样本数量大幅减少的情况下,应用线性核的模型检测准确率仍能达到100%。因此,结合MLSPTSVM模型的高光谱图像检测方法有效地提高热损伤大豆检测精度,且具有良好的鲁棒性,为大豆品质的检测提供了新的方法。

Detection of Heat-damaged Soybeans using Hyperspectral Imaging Technology and MLSPTSVM Model

In the process of transportation, imported soybeans are easy to cause heat damage due to high storage temperature. Heat damage aggravates the quality of soybean protein and oil and affects the quality of soybean. In this paper, hyperspectral imaging technology and multiple least squares recursive projection twin support vector machine (MLSPTSVM) were used to detect heat-damaged soybeans. Hyperspectral imaging detection system was used to obtain the spectral images of normal soybean, mild heat-damaged soybean and severe heat-damaged soybean in the range of 400~1000 nm. A variety of preprocessing methods were used for spectral pretreatment, and the effectiveness of the preprocessing methods for improving the model detection performance was analyzed. The experimental results show that both the MLSPTSVM model of a linear kernel with multiplicative scatter correction preprocessing and the MLSPTSVM model of a nonlinear kernel with the original spectral data can achieve 100% detection accuracy, which is much higher than other classical detection models. Moreover, the detection performance of the proposed model will not decrease when the number of experimental samples is significantly reduced. Therefore, the application of hyperspectral imaging technology combined with MLSPTSVM model can realize accurate, rapid and non-destructive detection of heat-damaged soybeans with good robustness, which provides a new method for soybean quality detection.