基于迁移学习的无监督细粒度图像分类模型  

Unsupervised Fine-grained Image Classification ModelBased on Transfer Learning

作　　者：韩天宇 张利锋[1] 王喜亮 HAN Tianyu;ZHANG Lifeng;WANG Xiliang(School of Information and Electrical Engineering,Ludong University,Yantai 264039,China;Stamping Workshop of South Factory,Shanghai General Motors Co.,Ltd.,Yantai 264006,China)

机构地区：[1]鲁东大学信息与电气工程学院,山东烟台264039 [2]上海通用汽车有限公司南厂冲压车间,山东烟台264006

出　　处：《鲁东大学学报（自然科学版）》2021年第2期139-145,共7页Journal of Ludong University:Natural Science Edition

摘　　要：图像分类是计算机视觉领域的一个重要研究分支,普通图像分类关注主类对象差异性的判别,而细粒度图像分类则重点研究主类下不同子类的区分。考虑到较小的子类间差异(只在某个局部上有细微差异)和由于拍摄角度、背景、姿态等因素导致的较大的子类内差异,细粒度图像识别成为一项颇具挑战的任务。为了缓解细粒度图像数据集难以获取和扩充的问题,本文提出了一种基于迁移学习和孪生网络实现的无监督细粒度图像分类模型。该模型采用两阶段训练策略,第一阶段采用有监督方式预训练基础网络,使模型快速学习到细粒度图像的通用特征;第二阶段通过搜索引擎扩充数据集,设计孪生网络并采用无监督方式微调模型以实现知识迁移。实验表明,孪生网络模型在Stanford Car和Aircraft数据集上能够获得较好的性能表现。Image classification is an important branch in computer vision applications.The common image classifications focus on discriminating the category of different main class objects,while the fine-grained image classifications concentrating on distinguishing the different sub-categories in a main class.The fine-grained image recognition tasks are quite challenging,considering the small differences between the subcategories(slightly differences of some specific parts of objects)and larger intra-subclass differences caused by shooting angles,background,postures and other factors.To alleviate the difficulty in acquiring and expanding of the fine-grained image datasets,an unsupervised fine-grained image classification model based on transfer learning in the form of twin networks was proposed.The model exploits a two-stage training strategy to train the twin network.The first stage pretrains the basic network by a supervised way so that the model can quickly learn the common features of the fine-grained images,while in the second stage,search engine is executed to expand the datasets and employ the twin network to fine-tunes it in an unsupervised manner in order to achieve the knowledge transferring.The experimental results show that the twin network model can achieve good performance on the Stanford Car and Aircraft datasets compared with the counterparts.

关 键 词：迁移学习 孪生网络 无监督学习 细粒度图像分类 

分 类 号：TP391.413[自动化与计算机技术—计算机应用技术]