植物图像识别方法研究及实现

发布时间：2018-01-10 03:23

本文关键词：植物图像识别方法研究及实现　出处：《浙江大学》2017年硕士论文　论文类型：学位论文

【摘要】：图像识别技术目前广泛应用于传统制造业、安保以及互联网行业,相关方法都较为成熟。但是,在生物、医学以及食品等领域上还有很多空白需要填补,主要在于现在的识别方法往往面向"大"类别识别,例如区分猫和狗、人和车辆、动物和植物等,并非细粒度(微差)图像识别的范畴,例如同为菊科下的秋菊和野菊的识别。本文针对微差图像识别领域中的植物图像识别进行了方法研究,主要工作如下:1.从图像底层特征入手,重点研究了 BoV和费雪向量特征编码方法,提出了基于费雪向量的多特征融合图像识别方案。实验表明,在植物图像识别应用中,基于费雪向量的特征编码方案具有更好的效果。2.从深度学习方法入手,首先设置对比实验进行模型初选,研究分析不同训练模式以及卷积神经网络深度对植物图像识别的影响;其次提出了基于选择性搜索算法的植物图像关键区域生成方法;最后提出了面向关键区域的基于VGGNet16的植物图像识别模型,并验证了本文提出方法的有效性。3.构建植物图像数据集。数据库的构建包含两部分,一是面向图像识别的公开数据集,用于方法的横向比较;二是自建的植物领域图像数据集,并在该数据集的基础上构建了常用药用植物图像集,用于验证方法的实用性。并将本文所提方法在数据集上进行实验。4.设计和实现药用植物图像识别系统。系统利用本文提出的具有较好效果的方法,在此基础上,设计了中间结果和最终结果的用户反馈机制,用以提高系统的图像识别准确率。
[Abstract]:Image recognition technology has been widely used in traditional manufacturing, security and Internet industries. However, there are still many gaps to be filled in biology, medicine and food. This is mainly due to the fact that current recognition methods are often oriented towards "large" category recognition, such as distinguishing between cats and dogs, humans and vehicles, animals and plants, and is not a category of fine-grained (micro-differential) image recognition. In this paper, the method of plant image recognition in the field of differential image recognition is studied. The main work is as follows: 1. Starting from the bottom features of the image. The method of BoV and Fisher vector feature coding is studied emphatically, and a multi-feature fusion image recognition scheme based on Fisher vector is proposed. The experimental results show that it is applied in plant image recognition. The feature coding scheme based on Fisher vector has a better effect. 2. Starting with the depth learning method, we first set up a comparative experiment to select the model. The effects of different training modes and the depth of convolution neural network on plant image recognition were studied. Secondly, based on the selective search algorithm, the key region generation method of plant image is proposed. Finally, a plant image recognition model based on VGGNet16 for key regions is proposed. The validity of the proposed method is verified. 3. The construction of plant image data set. The construction of database consists of two parts: one is the open data set for image recognition, which is used for the horizontal comparison of methods; Secondly, the image data set of plant domain was built, and the common medicinal plant image set was constructed on the basis of the data set. This method is used to verify the practicability of the method. The method proposed in this paper is tested on the data set. 4. The design and implementation of medicinal plant image recognition system. The system uses the method proposed in this paper with better results. On this basis, the user feedback mechanism of intermediate and final results is designed to improve the accuracy of image recognition.
【学位授予单位】：浙江大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】