高光谱成像结合SIS与RFS的蓝莓腐烂病检测

为了能够快速、无损地检测采后蓝莓腐烂病害,利用高光谱成像技术对采后蓝莓腐烂病进行检测。本研究提出光谱信息分割法(SIS)对866.5 nm波段图像蓝莓病害区域进行分割,取得了较好效果。另外根据正常蓝莓光谱曲线和病害蓝莓光谱曲线,提出区域特征筛选法(RFS),并结合竞争适应重加权采样(CARS)和连续投影(SPA)算法对可见光第一区域波段、近红外第二区域波段分别筛选特征波长。最后,采用相关向量机(RVM)模型和径向基(RBF)神经网络模型对蓝莓腐烂病害进行检测。检测结果表明,第一区域与第二区域特征波长组合建立的CARS-RBF模型检测效果最好,特征波长为655.8,710.9,752.2,759.9,761.2,866.5,969.7 nm,训练集和测试集正常蓝莓与病害蓝莓检测准确率分别为98.3%、98.6%和97.5%、98.75%。本研究提出的光谱信息分割法(SIS)和区域特征筛选法(RFS)检测蓝莓病害,为蓝莓在线检测和分拣提供了一种新的参考方法。

Detection of Rot Blueberry Disease by Hyperspectral Imaging with SIS and RFS

In order to detect the rot disease of postharvest blueberries quickly, effectively and accurately, the rot disease of postharvest blueberries was detected by hyperspectral imaging technology with different detection models. According to analyse the difference of the spectral relative reflectance between the normal blueberries and the disease blueberries, the spectral Information segmentation(SIS) was proposed to segment the disease regions of blueberries to solve the problem that the conventional threshold segmentation method is difficult to accurately segment blueberry disease regions due to the indistinct color characteristics of the normal blueberries regions and the disease blueberries regions. According to the difference of spectrum in 450 nm to 1 000 nm, the regional feature selection(RFS) was put forward that divided the spectral relative reflectance(450-1 000 nm) into the two regions. The first region was in visible spectrum ranges 450-780 nm and the second region in near infrared ranges 780-1 000 nm.Then CARS and SPA were used to extract the characteristic wavelengths from spectral data in two regions. Finally, relevance vector machines(RVM) model and radial basis function(RBF) model were used to detect the rot disease of blueberries. By comparing the detection effects of different models, the CARS-RBF model in the combined regions of first region and second region had best detection effect and the characteristic wavelengths were 655.8, 710.9, 752.2, 759.9, 761.2, 866.5, 969.7 nm. The detection accuracy of the normal blueberries and the disease blueberries in the training sets and the testing sets were 98.3%, 98.6% and 97.5%, 98.75%, respectively. According to the result of the detection,we can draw a conclusion that the spectral information segmentation(SIS) and regional feature selection(RFS) were used to detect blueberry diseases effectively, which provide a new reference method for on-line detection and sorting of blueberries.