查询结果:   许婷婷,朱允斌,张跃.基于联合分类器的非自然图像检索[J].计算机应用与软件,2018,35(4):244 - 248.
中文标题
基于联合分类器的非自然图像检索
发表栏目
图像处理与应用
摘要点击数
1327
英文标题
NON-NATRUAL IMAGE RETRIEVAL VIA COMBINED CLASSIFIER
作 者
许婷婷 朱允斌 张跃 Xu Tingting Zhu Yunbin Zhang Yue
作者单位
复旦大学计算机科学技术学院 上海 201203 上海视频技术与系统工程研究中心 上海 201203 华为技术有限公司 江苏 南京 210000   
英文单位
School of Computer Science,Fudan University,Shanghai 201203,China Shanghai Engineering Research Center for Video Technology and System,Shanghai 201203,China Huawei Technologies Co.,Ltd.,Nanjing 210000,Jiangsu,China   
关键词
非自然图像 联合分类器 卷积神经网络
Keywords
Non-natural image Combined classifier Convolutional neural network
基金项目
国家重点研发计划项目(2016YFC0801003);上海市科委科研计划项目(16511105402);上海市人才计划项目(17XD1425000)
作者资料
许婷婷,硕士生,主研领域:多媒体信息处理与检索。朱允斌,博士生。张跃,硕士生。 。
文章摘要
卷积神经网络已经被广泛应用于图像检索领域,目前的图像检索系统在自然图像上效果很好。但是自然世界中还存在着海报、漫画这样的非自然图像,这些图像的检索效果并不好。相比于自然图像,这些图像往往具有大面积的单一色块,相邻色块间经常存在强烈的颜色对比,色块的分布也更无规律,可提取特征较少。如果想要较好的检索效果,就需要提取较多的特征,进而需要设计更深、更复杂的网络,或者利用多个卷积神经网络的提取信息。尝试解决以海报图像为例的非自然图像检索,提出一个基于卷积神经网络的联合分类器。此联合分类器用不同的融合算法结合了多个独立分类器的结果,再进行分类。实验结果表明,对比单分类器,所提出的基于卷积神经网络的联合分类器能有效提高分类准确率。
Abstract
Convolutional neural network has been widely used in the field of image retrieval, and the current image retrieval system works well on natural images. But there are non-natural images such as posters, comic books in the natural world, and the retrieval system doesn’t work well on these images. Compared with natural images, non-natural images often have large areas of plain color patches, and there is often a strong color contrast between adjacent patches. Non-natural images usually have less extractable features. To get better retrieval results, more features are needed. So, it is necessary to design deeper, more complex networks, or extract information using multiple convolutional neural network. Trying to solve the non-natural image retrieval problem, this paper proposes a combined classifier based on convolutional neural network. The combined classifier incorporates multiple independent classifiers based on different fusion algorithms. The experimental results show that the method proposed in this paper can improve the classification accuracy effectively.
下载PDF全文