查询结果:   骆正茂.结合卷积神经网络不同层的特征进行包类商品检索[J].计算机应用与软件,2018,35(1):195 - 199,287.
中文标题
结合卷积神经网络不同层的特征进行包类商品检索
发表栏目
人工智能与识别
摘要点击数
810
英文标题
PACKAGE GOODS RETRIEVAL BASED ON UNITED DIFFERENT LAYERS OF CONVOLUTIONAL NEURAL NETWORK
作 者
骆正茂 Luo Zhengmao
作者单位
浙江东方职业技术学院 浙江 温州 325011     
英文单位
Zhejiang Dongfang Vocational and Technical College,Wenzhou 325011,Zhejiang,China     
关键词
图像检索 卷积神经网络 深度学习
Keywords
Image retrieval Convolutional neural network Deep learning
基金项目
浙江省教育厅2015年度高等学校国内访问学者专业发展项目(153)
作者资料
骆正茂,副教授,主研领域:图像处理,云计算技术。 [HJ2.35mm] 。
文章摘要
基于图像的内容相似性进行图像检索是计算机视觉领域的一个重要研究内容,比传统基于关键字的检索方法能节约相似物体的筛选时间。包类商品不仅市场需求量大且种类繁多,而且关键字往往不能准确地描述各种包类的款式。因此,利用关键字来进行包类款式搜索有时候并不能达到预期效果,基于内容图像检索则能较好地处理包类商品的检索问题。近年来,深度学习算法被用来解决图像相似搜索问题,但主要是基于卷积网络高层的特征作为输入图像的特征表达。高层的特征是比较抽象的语义特征,容易丢失一些图像的细节信息,因此会影响图像搜索的准确度。基于这个观察,运用基于卷积神经网络的深度学习算法完成对包类商品图像特征的自动学习,整理了5 000张分类好的包类商品图像作为深度学习的训练集。并且,和现有方法不同的是,结合了卷积神经网络中不同层的特征来提升包类商品图像检索的准确度。通过实验和实际应用验证了该方法能够较好地完成包类款式图像的检索问题。
Abstract
Image retrieval based on image similarity is an important research content in the field of computer vision, which can save the screening time of similar objects than traditional keyword-based retrieval methods. Packages have great marketing demand and contain lots of different kinds. However, it is hard to retrieve the same kind of package without accurate descriptions of keywords. Therefore, the use of keywords to carry out the package type search sometimes can’t achieve the desired results, based on the content image retrieval can be better to deal with package category of goods retrieval problems. In recent years, the depth learning algorithm has been used to solve the image similarity search problem, but mainly based on the characteristics of the high level of convolution network as the input image of the expression. High-level features are more abstract semantic features, easy to lose some of the details of the image information, it affect the accuracy of image search. Based on this fact, this paper used the convolutional neural network based deep learning method for the automatic learning of features of package images, and collected 5 000 classified package images as the training data. In addition, contrast to existing methods we combined the features of different layers in the convolutional neural network to improve the accuracy of package image retrieval. Experimental results and application showed that our algorithm was capable of retrieving package images.
下载PDF全文