节点文献

融合多层卷积神经网络特征的快速图像检索方法

A Fast Image Retrieval Method Based on Multi-Layer CNN Features

  • 推荐 CAJ下载
  • PDF下载
  • 不支持迅雷等下载工具,请取消加速工具后下载。

【作者】 王志明张航

【Author】 Wang Zhiming;Zhang Hang;School of Computer & Communication Engineering, University of Science and Technology Beijing;

【机构】 北京科技大学计算机与通信工程学院

【摘要】 基于卷积神经网络在图像特征表示方面的良好表现,以及深度哈希可以满足大规模图像检索对检索时间的要求,提出了一种结合卷积神经网络和深度哈希的图像检索方法.针对当前典型图像检索方法仅仅使用全连接层作为图像特征进行检索时,存在有些样本的检索准确率为零的问题,提出融合神经网络不同层的信息作为图像的特征表示;针对直接使用图像特征进行检索时响应时间过长的问题,使用深度哈希的方法将图像特征映射为二进制的哈希码,这样哈希码中既包含底层的边缘信息又包含高层的语义信息;同时,提出了一种相似性度量函数进行相似性匹配.实验结果表明,与已有的图像检索方法相比,该方法在检索准确率上有一定程度的提高.

【Abstract】 Based on the excellent performance of convolutional neural network in image feature representation, and deep hashing can meet the retrieval time requirement of large-scale image retrieval, this paper proposes an image retrieval algorithm combining convolutional neural network and deep hashing. The typical image retrieval algorithm only uses the full connection layer as the feature of image retrieval, and some of the samples have the retrieval accuracy of 0. We propose to fuse the information of different layers of a neural network as the feature representation of an image. Aiming at the problem of long response time when directly using image features for retrieval, we propose to use deep hashing to map the image features into binary hash codes, so that hash codes contain both the low-level edge information and high-level semantic information. Meanwhile, we propose a new similarity measure function for similarity matching.Compared with the existing image retrieval algorithms, experimental results show that our algorithm makes some improvements in retrieval accuracy.

  • 【文献出处】 计算机辅助设计与图形学学报 ,Journal of Computer-Aided Design & Computer Graphics , 编辑部邮箱 ,2019年08期
  • 【分类号】TP391.41;TP183
  • 【被引频次】23
  • 【下载频次】566
节点文献中: 

本文链接的文献网络图示:

本文的引文网络