基于卷积神经网络的多标签图像自动标注  

Multi-label Image Annotation Based on Convolutional Neural Network

作　　者：黎健成 袁春[1] 宋友[2] 

机构地区：[1]清华大学计算机科学与技术系,北京100084 [2]北京航空航天大学软件学院,北京100191

出　　处：《计算机科学》2016年第7期41-45,共5页Computer Science

基　　金：国家自然科学基金(U1433112;61170253);国家核高基项目(2013ZX01039001-002);国家高技术研究发展计划("863计划")(2011AA01A205);清华-腾讯合作项目(人体虚拟形象建模)资助

摘　　要：如今生活中,图像资源无处不在,海量的图像让人应接不暇。如何快速有效地对这些图像信息进行查询、检索和组织,成为了当前亟需解决的热门问题。而图像自动标注是解决基于文本的图像检索的关键。文中提出的这套基于深度学习模型中的卷积神经网络模型的多标签图像自动标注系统,实现了多标签损失排名函数,完成了多标签数据的训练与测试。在实验验证上,先选取CIFAR-10数据集进行算法的有效性测试,然后选取多标签图像数据集Corel 5k进行定量测试比较,结果表明,该算法的综合性能指标与现有算法相比有较大的提升。In today＇s life, the image resource is almost ubiquitous. An ocean ot tmages make peopLe overwnelmed. How to query, retrieve and organize these image information quickly and effectively is an urgent hot issue. The automatic image annotation is the key of text-based image retrieval solutional. A multi-label image annotation system based on a well- known deep learning model, convolutional neural network, was proposed in this paper, together with a multi-label loss ranking function to complete, the training and testing of multi-label image dataset. In the experiments, firstly, CIFAR-10 dataset were selected to test the effectiveness of the algorithm, and then quantitative test comparsion was conducted on multi-label image dataset Corel 5k. The proposed solution shows superior performance over the conventional algorithm.

关 键 词：图像自动标注 多标签 深度学习 卷积神经网络 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]