基于深度学习的植物图像集识别技术研究
发布时间：2018-11-08 18:43
【摘要】：植物分类学是一门对植物物种进行准确描述,命名,分群归类,探求各类群之间的亲缘关系,以及演化过程的基础科学。随着模式识别技术在植物图像分类任务中的广泛应用,对植物分类学的发展具有促进作用,而且给农业的科学研究带来了非常大的帮助。相比于传统的基于单个或少量图像的植物图像分类识别算法,基于图像集的植物图像分类算法的关键是如何为图像集进行建模,以及如何度量对图像集建模后模型之间的相似度。为了更好地对植物图像进行分类识别,有必要对以植物叶片图像集为研究对象的分类技术进行研究。本文以植物叶片图像为研究目标,对非线性重构模型、SPCANet模型、KmeansNet模型和利用深度模型对植物图像进行粒度分类等内容进行详细的介绍。本文的主要工作内容为:(1)提出一种基于非线性重构模型的植物叶片图像集的分类识别方法。该方法使用高斯受限玻尔兹曼机(GRBMs)通过非监督预训练来初始化模型的权值,然后为每一个植物叶片图像集用初始化的模型训练得到一个特定的模型。最后根据测试样本的最小重构误差和测试样本集的最多投票策略来判定测试样本集的类别。并采用基于k-means的特征提取方法来提取植物叶片图像特征。(2)提出了一种浅层PCANet(SPCANet)模型的植物图像集的分类识别方法。该方法首先用SPCANet模型来提取植物图像的特征,然后用线性SVM分类,最后根据投票策略判定测试集的类别。该模型是基于卷积神经网的结构设计的。该模型由卷积滤波层、非线性层和特征提取层三部分组成,其中卷积层的卷积核不同于传统的深度学习网络,而是通过PCA算法得到,这大大的减少了网络的训练时间和参数的设置。(3)提出一种KmeansNet模型的植物图像集的分类识别方法。该方法是SPCANet模型的变体,不同之处在于卷积层的卷积核是通过Kmeans算法得到。(4)利用深度学习Caffe框架对大规模植物图像进行粒度分类。引入粒度分类的思想为大规模植物图像的分类提供了一个新的思路。在大数据的背景下,利用Caffenet模型强大规模分类能力,通过微调Caffenet网络以实现对植物图像分别按门、纲、目、科、属进行粒度分类。
[Abstract]:Plant taxonomy is a basic science for the accurate description, naming, grouping and classification of plant species, exploring the relationship between various groups, and the evolution process. With the wide application of pattern recognition technology in the task of plant image classification, it can promote the development of plant taxonomy and bring great help to the scientific research of agriculture. Compared with the traditional plant image classification and recognition algorithm based on a single or small number of images, the key of the plant image classification algorithm based on image set is how to model the image set. And how to measure the similarity between the models after modeling the image sets. In order to better classify and recognize plant images, it is necessary to study the classification technology of plant leaf image set. In this paper, the nonlinear reconstruction model, SPCANet model, KmeansNet model and granularity classification of plant image by depth model are introduced in detail. The main work of this paper is as follows: (1) A classification and recognition method of plant leaf image set based on nonlinear reconstruction model is proposed. The method uses Gao Si constrained Boltzmann machine (GRBMs) to initialize the weight of the model by unsupervised pre-training, and then trains a specific model for each plant leaf image set with the initialized model. Finally, according to the minimum reconstruction error of the test sample and the maximum voting strategy of the test sample set, the classification of the test sample set is determined. The feature extraction method based on k-means is used to extract the feature of plant leaf image. (2) A classification and recognition method of plant image set based on shallow PCANet (SPCANet) model is proposed. Firstly, the SPCANet model is used to extract the features of plant images, then the linear SVM classification is used. Finally, the classification of the test set is determined according to the voting strategy. The model is based on the structure of the convolutional neural network. The model consists of three parts: convolution filter layer, nonlinear layer and feature extraction layer. The convolution kernel of the convolution layer is different from the traditional depth learning network, but is obtained by PCA algorithm. This greatly reduces the training time and parameter setting of the network. (3) A classification and recognition method of plant image set based on KmeansNet model is proposed. This method is a variant of SPCANet model, the difference is that the convolution kernel of convolution layer is obtained by Kmeans algorithm. (4) granularity classification of large scale plant images is carried out by using depth learning Caffe framework. The idea of granularity classification is introduced to provide a new idea for the classification of large-scale plant images. Under the background of big data, the Caffenet model is used to classify plant images according to door, class, order, family and genus by fine-tuning Caffenet network.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2319335
[Abstract]:Plant taxonomy is a basic science for the accurate description, naming, grouping and classification of plant species, exploring the relationship between various groups, and the evolution process. With the wide application of pattern recognition technology in the task of plant image classification, it can promote the development of plant taxonomy and bring great help to the scientific research of agriculture. Compared with the traditional plant image classification and recognition algorithm based on a single or small number of images, the key of the plant image classification algorithm based on image set is how to model the image set. And how to measure the similarity between the models after modeling the image sets. In order to better classify and recognize plant images, it is necessary to study the classification technology of plant leaf image set. In this paper, the nonlinear reconstruction model, SPCANet model, KmeansNet model and granularity classification of plant image by depth model are introduced in detail. The main work of this paper is as follows: (1) A classification and recognition method of plant leaf image set based on nonlinear reconstruction model is proposed. The method uses Gao Si constrained Boltzmann machine (GRBMs) to initialize the weight of the model by unsupervised pre-training, and then trains a specific model for each plant leaf image set with the initialized model. Finally, according to the minimum reconstruction error of the test sample and the maximum voting strategy of the test sample set, the classification of the test sample set is determined. The feature extraction method based on k-means is used to extract the feature of plant leaf image. (2) A classification and recognition method of plant image set based on shallow PCANet (SPCANet) model is proposed. Firstly, the SPCANet model is used to extract the features of plant images, then the linear SVM classification is used. Finally, the classification of the test set is determined according to the voting strategy. The model is based on the structure of the convolutional neural network. The model consists of three parts: convolution filter layer, nonlinear layer and feature extraction layer. The convolution kernel of the convolution layer is different from the traditional depth learning network, but is obtained by PCA algorithm. This greatly reduces the training time and parameter setting of the network. (3) A classification and recognition method of plant image set based on KmeansNet model is proposed. This method is a variant of SPCANet model, the difference is that the convolution kernel of convolution layer is obtained by Kmeans algorithm. (4) granularity classification of large scale plant images is carried out by using depth learning Caffe framework. The idea of granularity classification is introduced to provide a new idea for the classification of large-scale plant images. Under the background of big data, the Caffenet model is used to classify plant images according to door, class, order, family and genus by fine-tuning Caffenet network.
【学位授予单位】：华侨大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前3条
1 龚丁禧;曹长荣;;基于卷积神经网络的植物叶片分类[J];计算机与现代化;2014年04期
2 翟传敏;汪青萍;杜吉祥;;基于叶缘与叶脉分数维特征的植物叶识别方法研究[J];计算机科学;2014年02期
3 杜吉祥;汪增福;;基于径向基概率神经网络的植物叶片自动识别方法[J];模式识别与人工智能;2008年02期
，本文编号：2319335
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2319335.html
