高光谱成像技术的库尔勒梨早期损伤可视化检测研究

发布时间：2018-02-24 01:01

本文关键词： 高光谱成像损伤检测支持向量机波段比运算库尔勒梨　出处：《光谱学与光谱分析》2017年01期 　论文类型：期刊论文

【摘要】：利用高光谱成像技术对库尔勒梨早期损伤进行快速识别检测。以60个库尔勒梨为研究对象,采集380~1 030nm波段范围内完好样本和损伤后1~7天样本的480幅高光谱图像。提取图像中感兴趣区域(ROI)的平均光谱信息,利用小波变换(WT)对光谱数据进行去噪平滑,将去噪后的全部样本按2∶1的比例分成建模集(320个)和预测集(160个)。利用二阶导数从全谱信息中提取出19个特征波长,分别基于全谱和提取出的特征波长对建模集和预测集进行支持向量机(SVM)建模分析。结果表明,基于全谱和特征波长的判别分析模型中,两者预测集的识别率都达到93.75%,表明提取的特征波长包含了光谱数据中的关键信息。然后,基于特征波长运用波段比运算挑选最佳波段比,根据波段比F值的分布确定光谱图像分割的最佳波长684和798nm。对最佳波段比(684/798nm)下的图像,利用选择性搜索(SS)对高光谱图像中样本的完好和损伤区域进行分割,从分割结果来看,1~7天损伤样本的受损区域能够被准确检测出来。研究结果表明:基于高光谱成像技术对库尔勒梨进行损伤鉴别是可行的,该研究所获得的特征波长和波段比为研发在线实时的库尔勒梨损伤检测系统提供支撑。
[Abstract]:Using hyperspectral imaging technology of Korla pear injury at the early stage of fast recognition detection. In 60 Korla pear as the research object, collecting 380~1 030nm band within 1~7 days after injury and intact samples of 480 hyperspectral images. The image region of interest extraction (ROI) the average spectral information, using wavelet transform (WT) the spectral data denoising smoothing, all samples will go after noise by 2: 1 ratio into modeling set (320) and prediction set (160). By using the two order derivative from the full spectrum of extracted 19 characteristic wavelengths based on full spectrum information, and extract the characteristic wavelength of the calibration set and the prediction set for support vector machine (SVM) modeling and analysis. The results show that the model based on discriminant analysis and full spectrum characteristic wavelength, the prediction of recognition rate has reached 93.75%, that includes the extraction of characteristic wavelength spectral data in the key Information. Then, using the characteristic wavelength band operation than the selection of the best band ratio based on band ratio distribution according to F-measure determine the optimal wavelength spectral image segmentation and 798nm. 684 of the best band ratio (684/798nm) under the image, using the selective search (SS) samples of hyperspectral images in the intact and damage region segmentation. From the segmentation results, 1~7 days damage samples of the damaged area can be accurately detected. The results show that: Based on hyperspectral imaging technology for damage identification of Korla pear is feasible, the study obtained characteristic wavelength and band ratio of Korla pear damage detection system for the development of online real-time support.

【作者单位】：浙江大学生物系统工程与食品科学学院;浙江大学农业与生物技术学院;
【基金】：南方山地果园智能化管理技术与装备协同创新中心开放基金项目(JX2014XCH103) 国家自然科学基金项目(31471417) 国家支撑计划项目(2015BAD19B03)资助
【分类号】：TP391.41;S661.2

【相似文献】