基于高光谱成像的苹果病害检测识别方法的研究

发布时间：2018-05-26 01:30

  本文选题：苹果病害检测 + 高光谱成像　； 参考：《沈阳农业大学》2017年硕士论文

【摘要】：苹果是我国水果市场上最常见的水果之一,但在其生长和存储过程中易受病害影响,从而会造成大量的经济损失。因此病害果的前期分拣十分必要。目前苹果病害的检测主要以人工分拣为主,但其分拣难度较大、准确性差、很难达到分级的一致性。因此本研究采用高光谱成像技术对苹果的病害进行快速、无损检测,其成果对提高苹果品质检测和分级水平具有重要意义。本研究主要研究内容及成果:(1)本研究以北方大面积种植的寒富苹果为研究对象,经调研发现寒富苹果常见的病害有炭疽病、苦痘病、褐斑病和黑腐病。为了提取少量的特征波长对苹果病害进行检测,利用改进流行距离法(IMD)、马氏距离法(MD)和连续投影法(SPA)3种方法提取特征波长。通过对比发现本研究所提出的二次连续投影算法SPA2提取3个特征波长(681、867和942nm)对苹果病害检测效果最佳。(2)本研究采用正常苹果和病害苹果的感兴趣区域的纹理特征或SPA2提取的3个特征波长光谱相对反射率的光谱特征作为特征向量,分别建立线性判别分析(LDA)、支持向量机(SVM)和BP神经网络(BP)模型检测苹果病害,得出SPA2-BP为苹果病害的最佳检测方法,训练集检测正确率达到100%,验证集检测正确率为98%。试验结果表明,利用少量的光谱信息采用BP神经网络可以有效地对苹果病害进行检测,能正确检测苹果是否侵染了病害。(3)本研究通过提取的感兴趣区域分割出图像进行纹理特征提取,及采用SPA2提取的3个有效波长的光谱相对反射率作为光谱特征形成三个不同的特征向量组合。利用这些特征向量组合构建BP神经网络和支持向量机模型对4种苹果病害进行识别,得出光谱特征结合纹理特征作为输入矢量的SVM检测模型对苹果病害识别效果最佳。验证集中正常果的检测正确率为95%,炭疽病的检测效果稍差为90%,苦痘病的检测正确率为95%,褐斑病的检测正确率95%,黑腐病的检测正确率为95%。试验结果表明,光谱特征结合纹理特征采用支持向量机建立识别模型可以有效地对苹果病害进行分类检测,为构建多光谱在线检测水果品质分级提供了理论依据。
[Abstract]:Apple is one of the most common fruits in Chinese fruit market, but it is easy to be affected by diseases in its growth and storage process, which will cause a lot of economic losses. Therefore, it is very necessary for the early stage sorting of the diseased fruit. At present, the main detection of apple diseases is manual sorting, but its sorting is difficult and accurate, so it is difficult to achieve the consistency of classification. Therefore, hyperspectral imaging technology is used to detect apple diseases quickly and nondestructive, and the results are of great significance to improve apple quality detection and grading level. The main contents and results of this study were as follows: (1) in this study, the common diseases of cold rich apple were anthracnose, bitter acne, brown spot and black rot, which were planted in a large area in the north of China. The common diseases of cold rich apple were anthracnose, bitter acne, brown spot and black rot. In order to extract a small amount of characteristic wavelengths for detection of apple diseases, the improved popular distance method (IMD), the Markov distance method (MDD) and the continuous projection method (spa) were used to extract the characteristic wavelengths. It is found that the quadratic continuous projection algorithm (SPA2) proposed in this study has the best effect on apple disease detection by extracting three feature wavelengths (681867 and 942nm).) the texture features of normal apple and diseased apple region of interest are used in this study. Or the spectral features of the relative reflectance of the three characteristic wavelengths extracted by SPA2 as feature vectors, The models of linear discriminant analysis (LDA), support vector machine (SVM) and BP neural network (BP) were established to detect apple diseases. The results showed that SPA2-BP was the best detection method for apple diseases. The correct rate of training set was 100 and the correct rate of verification set was 98. The experimental results show that BP neural network can be used to detect apple diseases effectively by using a small amount of spectral information. In this study, the region of interest was extracted and the image was segmented for texture feature extraction. The spectral relative reflectance of three effective wavelengths extracted by SPA2 is used as the spectral feature to form three different eigenvector combinations. BP neural network and support vector machine (SVM) model are used to identify four apple diseases. It is concluded that the SVM detection model with spectral features and texture features as input vectors is the best for apple disease recognition. The accuracy rate of testing normal fruit was 95%, that of anthracnose was 90%, that of bitter acne was 95%, that of brown spot was 95%, and that of black rot was 95%. The experimental results show that the recognition model based on support vector machine (SVM) combined with spectral features and texture features can effectively classify and detect apple diseases and provide a theoretical basis for the construction of multispectral on-line fruit quality classification.
【学位授予单位】：沈阳农业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;S436.611
，

本文编号：1935420