单位代码 10476
学号 **********
igrtTP391纳吉
硕士学位论文
（专业学位）
东风金龙专业学位领域：农业信息化
专业学位类别：农业硕士
***：郜翔
指导教师：刘尚旺 副教授
二〇一八年四月
FLOWER IMAGE CLASSIFICATION BASED ON DEEP MODEL TRANSFER
LEARNING
远程教育论坛
A Dissertation Submitted
to the Graduate School of Henan Normal University in Partial Fulfillment of the Requirements
for the Degree of Master of Agricultural
By
Gao Xiang
Supervisor：Associate Prof. Liu Shangwang
April，2018
摘 要
图像分类是目前人工智能、模式识别领域的重要研究方向之一。花卉图像分类，是以花卉为目标，对图像中的花卉进行生物学分类；其在植物物种研究与保护、园林花圃智能化管理中具有重要应用价值。花卉图像类别数量多、类间差异小、类内差异大、背景复杂、样本数量少，对其准确分类往往难度较大。针对花卉图像数据的这些特点，本文以深度模型、迁移学习和多任务学习为理论指导，进行花卉图像分类研究。主要工作内容如下： （1）针对特征描述算子设计难度大、特征提取能力弱，以及深度模型结构复杂、参数规模大、难以拟合小数据集的问题，提出基于深度模型迁移学习的细粒度图像分类方法。首先，通过在粗粒度图像数据集上进行预训练，使深度模型参数分布具备自然图像特征提取能力；然后，对深度模型进行局部训练，使其在细粒度图像数据集上进行迁移。实验结果表明，在102类花卉图像数据集上，分类准确率达到96.27%；此外，在120类狗、200类鸟、37类猫和狗、196类汽车图像数据集上，该方法亦分别得到72.23%、73.33%、86.00%、89.72%的图像分类准确率，具有较好的准确性与泛化性能。经历我的1957年
（2）深度模型的训练往往需要非常大的数据规模、非常多的差异特征，以满足模型中庞大的参数规模要求；然而，细粒度图像数据集规模相比粗粒度图像数据集规模要小得多。虽然对原始数据进行不同方式的修改，可以实现数据增广，但引入的差异特征不够明显，因而性能提升有限。为此，提出基于网络爬虫的外部数据增广方法。首先，在社交平台、搜索引擎上（例如Instagram、Flickr、Google、Bing、Baidu），根据Hashtag、Keyword的搜索结果，无差别爬取图像（难免包含大量冗余图像）；然后，对冗余图像进行筛选、整理，生成增广数据集；最后，使用增广数据集取代原始数据集，对模型进行迁移训练。实验结果显示，在增广后的102类花卉图像数据集上，图像分类准确率达到99.41%，较原始数据集提高3.14%。相比于原始数据集，使用外部增广数据能够有效提高分类准确率。
（3）根据实际需求，设计、实现基于B/S架构的花卉图像分类原型系统。首先，将迁移模型导出、参数冻结、线上部署；再通过浏览器远程调用服务器上的模型进行图像分类，并在WEB端给出分类结果。在6种移动终端上进行测试，结果显示，在共计
18幅花卉图像中，仅有2幅结果为Top-2，其余均为Top-1；在不考虑网络延迟的情况下，服务器响应速度控制在40ms以内。证明本系统具有较高准确性与即时性。
关键词：深度模型，迁移学习，图像分类，多任务学习，数据增广，网络爬虫
川端康成
ABSTRACT
Image classification is one of the important research directions in the field of artificial intelligence and pattern recognition. Flower image classification is to classify the main flowers in images. It has important application value in research and protection of plant species and intelligent management of garden flowers and plants. Because of the huge number of categories, small differences between classes, large differences within the category, the complex background, and the small number of samples, it is often difficult to classify flowers accurately. Based on the characteristics of flower image data and the existing problems of classification algorithm, guided by the theory of deep learning, transfer learning, and multitask learning, flower image classification is studied in this thesis. The main work is as follows:
（1）Aiming at overcoming difficulties of designing feature description operator, weak ability of feature extraction; and solving the problems of complex structure, large parameter scale, difficult fitting small datasets of deep model, a fine grained image classification method of deep learning integrated with transfer learning is proposed. Firstly, after pre-training on the coarse-grained image dataset, the parameter distribution of deep model was able to extract natural image features. Secondly, the deep model is locally trained to transfer on fine-grained image datasets. The experime
ntal results show that the classification accuracy rate reaches 96.27% on 102 kinds of flower image datasets. The method also obtained 72.23%, 73.33%, 86.00%, and 89.72% classification accuracy on the 120 class dog, 200 class bird, 37 class cat and dog, and 196 class car image datasets, respectively.
（2）Deep model training often requires a very large dataset size, a large number of differences in characteristics to meet the large parameter scale in the model, and the fine-grained image dataset scale is much smaller than the coarse-grained image dataset size. Modifying the original data in different ways can increase the data, but the introduced difference features are not obvious. In response to this problem, this paper proposes an external data augmentation method based on WEB crawlers. Firstly, indiscriminately crawl images from the social platforms, search engines (such as Instagram, Flickr, Google, Bing, Baidu) according to Hashtag, Keyword search results which inevitably contains a large number of redundant images. Then, the redundant images are filtered and organized to generate augmented datasets. Finally, use
豫公网安备41160202000603
站长QQ:729038198
关于我们
投诉建议