基于深度信念网络的植物叶片识别研究

发布时间：2019-04-10 08:49

【摘要】：在尽量短的时间内实现叶片图像识别对研究、保护植物具有重要的实践意义；然而人工分析和处理爆炸性增长的叶片图像数据几乎是不可能完成的。因而,使用计算机进行辅助的植物识别来提高植物叶片图像识别效率、减少人工成本是现在的研究热点。传统的叶片识别由于提取的特征不够充分或者是分类器模型复杂度不够,导致叶片种类增加到一定数量后识别率不够高。针对这个问题,本文提出使用多特征融合的特征向量和深度信念网络分类器进行叶片识别。本文的主要研究内容和结论如下：1、本文使用典型的Flavia叶片数据库和ICL数据库进行实验。首先我们对叶片图像进行预处理,包括滤波、旋转、剪裁等,去除图像拍摄问题对识别的干扰。然后计算了傅里叶描述子、Gabor滤波特征、局部二值模式、Hu不变量、灰度共生矩阵特征构成叶片特征向量。2、我们使用能够建立复杂的分类构架的深度信念网络作为分类器。在分类器训练过程中,使用"dropout"算法减少其过拟合现象的产生。这个算法对Flavia数据库中的32种叶片和ICL数据库中的220种叶片的识别率分别为99.37%、91.2%,并与其他研究者的结果进行了比较。我们还通过实验分析训练样本数量、隐含层单元数、批训练大小等因素对深度信念网络识别效果的影响,为设计良好的网络结构提供依据。3、由于预训练过程对深度信念网络的影响很大,所以本文提出了2种改进的预训练算法以进一步提高识别率。第1种称为Mean-DBNs算法,采用参数取其训练平均值作为预训练步骤的结果,在ICL数据库中对220种叶片的识别率为93.1%。第2种是在参数调节的过程中增加了积分环节和微分环节,称作PID-DBNs算法。实验显示这种改进算法对220种叶片的识别率可以达到94.1%,并且比原来的网络所需要的训练时间更短、稳定性更强。
[Abstract]:The realization of leaf image recognition in the shortest possible time is of great practical significance to the research and protection of plants, however, it is almost impossible to analyze and process the explosively growing leaf image data manually. Therefore, the use of computer-aided plant recognition to improve the efficiency of plant leaf image recognition, reduce labor costs is now the focus of research. Because of the insufficient extracted features or the insufficient complexity of classifier model, the recognition rate of the traditional leaf recognition is not high enough when the number of leaves is increased to a certain number. In order to solve this problem, this paper proposes a multi-feature fusion feature vector and depth belief network classifier for blade recognition. The main contents and conclusions of this paper are as follows: 1. The typical Flavia blade database and ICL database are used to experiment in this paper. First, we pre-process the blade image, including filtering, rotation, clipping and so on, to remove the interference of the image shooting problem to the recognition. Then Fourier descriptors, Gabor filtering features, local binary patterns, Hu invariants, gray-level co-occurrence matrix features are calculated to form the blade feature vector. 2. We use the depth belief network which can build complex classification framework as the classifier. In the process of classifier training, "dropout" algorithm is used to reduce the occurrence of over-fitting. The recognition rates of 32 blades in Flavia database and 91.2% in ICL database are 99.37% and 91.2% respectively. The results are compared with those of other researchers. We also analyze the influence of the number of training samples, the number of hidden layer units and the size of batch training on the recognition effect of depth belief network through experiments, which provide a basis for designing a good network structure. 3, Because the pre-training process has a great influence on the depth belief network, this paper proposes two improved pre-training algorithms to further improve the recognition rate. The first one is called Mean-DBNs algorithm. The training average value of the parameters is used as the result of the pre-training step. The recognition rate of 220 kinds of blades in ICL database is 93.1%. The second is to add integral and differential links in the process of parameter adjustment, which is called PID-DBNs algorithm. The experiments show that the recognition rate of 220 blades can reach 94.1%, and the training time is shorter and the stability is stronger than the original network.
【学位授予单位】：北京林业大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】