基于流形学习的葡萄叶片品种识别方法研究

发布时间：2018-08-13 10:42

【摘要】：随着葡萄市场经济的发展,葡萄品种识别对于葡萄这种经济作物的科普和市场推广都具有重要意义。在葡萄品种识别研究中,一般以叶片为识别研究对象,主要考虑到叶片相对于果实来说易于保存且可采摘时间长,研究过程中不需要其他学科的辅助实验。但是葡萄叶片识别存在一个明显的难点,这种同科属类叶片颜色和形态结构差异小,使得识别研究中准确率不高。为解决这个问题,本论文展开了针对同属类葡萄叶片识别的研究,提出一种基于流形学习的葡萄叶片品种识别方案。本研究采用了15个品种的450幅葡萄叶片作为实验样本,进行葡萄品种的分类识别实验。重点研究葡萄叶片的特征提取和特征降维。在特征提取部分,分别采用人工设计特征的灰度共生矩阵、方向梯度直方图、可变性部件模型和由卷积神经网络提取的深度学习特征作为叶片特征,分析特征数据性质和表现能力。发现高维度的特征表示葡萄叶片的能力优于低维度,但是高维特征数据量大,冗余性高,虽然能够得到较好的识别结果,但效率较低。为了降低高维葡萄叶片特征数据在识别过程中的复杂度,提高其实用性与实验效率,本文采用流形学习算法对提取的高维葡萄叶片特征进行降维,在保持识别精度的基础上,提高算法的效率,使其具有实用性。在葡萄叶片特征降维研究中,分别应用了局部线性嵌入(LLE)、拉普拉斯特征映射(LE)、局部保持投影(LPP)、临近保持嵌入(NPE)四种不同的算法进行特征降维,得到葡萄叶片在低维空间的特征表示,并对影响降维性能的重点参数进行了分析;在葡萄叶片识别过程中,对比分析不同分类器的分类效果,最后通过训练支持向量机(SVM)分类模型,进行叶片分类识别。本论文通过实验分析验证了流形降维在葡萄叶片识别中的可行性和必要性,流形降维能够有效保持数据在高维空间的内部结构特征。降维后的特征,在提高识别速度的同时,相对于降维前,仍具有良好的叶片识别准确性。其中利用卷积神经网络进行特征提取结合流形学习算法对其进行降维,识别率最高可以达到90.33%,识别性能优于降维前的性能,并且识别速度大幅提高,相较于不做特征降维的识别时间而言,时间缩短为原来的1/3。对于人工设计特征的降维,降维后的识别时间有明显的改善,其中DPM特征维度降为原来的1/30时,识别用时缩短为原来的1/6。本文的研究为葡萄叶片的快速识别,提供了一种有效的方法。
[Abstract]:With the development of grape market economy, grape variety identification is of great significance to the popularization of science and marketing of this kind of cash crop. In the research of grape variety recognition, the leaf is generally regarded as the object of study, considering that the leaf is easy to be preserved and can be picked for a long time compared with the fruit, and there is no need for the auxiliary experiment of other disciplines in the course of the research. However, there is an obvious difficulty in the leaf recognition of grape. The difference of leaf color and morphological structure of the same family, genus and species is small, which makes the accuracy of the recognition research not high. In order to solve this problem, this paper studies the leaf recognition of the same genus grape, and proposes a new method of grape leaf variety recognition based on manifold learning. In this study, 450 leaves of 15 grape varieties were used as experimental samples to classify and identify grape varieties. The feature extraction and dimensionality reduction of grape leaves were studied. In the part of feature extraction, the grayscale co-occurrence matrix, directional gradient histogram, variable component model and depth learning feature extracted from convolutional neural network are used as leaf features, respectively. Analyze the nature and performance of the feature data. It was found that the high dimension feature indicates that the ability of grape leaves is superior to that of the low dimension, but the high dimensional feature has large data volume and high redundancy, which can obtain better recognition results, but its efficiency is low. In order to reduce the complexity of high dimensional grape leaf feature data in recognition process and improve its practicability and experimental efficiency, this paper adopts manifold learning algorithm to reduce the dimension of the extracted high dimensional grape leaf feature, on the basis of maintaining the recognition accuracy. Improve the efficiency of the algorithm and make it practical. In the study of grape leaf feature dimensionality reduction, four different algorithms of locally linear embedded (LLE), Laplacian feature map, (LE), local preserving projection, (LPP), near preserving embedding (NPE), were used to reduce the dimension of grape leaves. The characteristic representation of grape leaves in low-dimensional space was obtained, and the key parameters affecting the performance of reducing dimension were analyzed. In the process of grape leaf recognition, the classification effect of different classifiers was compared and analyzed. Finally, by training support vector machine (SVM) classification model, the blade classification recognition. In this paper, the feasibility and necessity of manifold dimensionality reduction in grape blade recognition are verified by experimental analysis. Manifold dimensionality reduction can effectively maintain the internal structural characteristics of data in high-dimensional space. The features after dimensionality reduction can improve the recognition speed and still have good accuracy of blade recognition compared with those before dimensionality reduction. Among them, convolution neural network is used for feature extraction and manifold learning algorithm to reduce the dimension, the recognition rate can reach 90.33, the recognition performance is better than that before dimensionality reduction, and the recognition speed is greatly improved. Compared with the recognition time without feature reduction, the time is shortened to one third of the original time. For the dimensionality reduction of artificial design features, the recognition time after dimensionality reduction is obviously improved. When the DPM feature dimension is reduced to 1 / 30 of the original, the recognition time is shortened to 1 / 6 of the original. The research in this paper provides an effective method for rapid recognition of grape leaves.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：S663.1;TP391.41

【参考文献】