基于深度卷积神经网络和二进制哈希学习的图像检索方法

发布时间：2018-05-18 10:20

本文选题：图像检索 + 深度卷积神径网络　；参考：《电子与信息学报》2016年08期

【摘要】：随着图像数据的迅猛增长,当前主流的图像检索方法采用的视觉特征编码步骤固定,缺少学习能力,导致其图像表达能力不强,而且视觉特征维数较高,严重制约了其图像检索性能。针对这些问题,该文提出一种基于深度卷积神径网络学习二进制哈希编码的方法,用于大规模的图像检索。该文的基本思想是在深度学习框架中增加一个哈希层,同时学习图像特征和哈希函数,且哈希函数满足独立性和量化误差最小的约束。首先,利用卷积神经网络强大的学习能力挖掘训练图像的内在隐含关系,提取图像深层特征,增强图像特征的区分性和表达能力。然后,将图像特征输入到哈希层,学习哈希函数使得哈希层输出的二进制哈希码分类误差和量化误差最小,且满足独立性约束。最后,给定输入图像通过该框架的哈希层得到相应的哈希码,从而可以在低维汉明空间中完成对大规模图像数据的有效检索。在3个常用数据集上的实验结果表明,利用所提方法得到哈希码,其图像检索性能优于当前主流方法。
[Abstract]:With the rapid growth of image data, the current mainstream image retrieval methods adopt fixed coding steps of visual features, lack of learning ability, resulting in poor ability of image expression and high dimension of visual features. The performance of image retrieval is seriously restricted. In order to solve these problems, this paper proposes a new method for learning binary hash coding based on deep convolutional shrink-path network, which can be used in large-scale image retrieval. The basic idea of this paper is to add a hash layer to the depth learning framework and to study image features and hash functions at the same time, and the hash functions satisfy the constraints of independence and minimum quantization error. Firstly, using the powerful learning ability of the convolution neural network to mine the implicit relation of the training image, extract the deep feature of the image, and enhance the distinguishing and expressing ability of the image feature. Then, the feature of the image is input to the hash layer, learning the hash function to minimize the classification error and quantization error of the binary hash code output from the hash layer, and satisfy the independence constraint. Finally, a given input image is obtained by the hash layer of the frame, so that the large scale image data can be effectively retrieved in low-dimensional hamming space. The experimental results on three common data sets show that the proposed method is superior to the current mainstream method in image retrieval performance.
【作者单位】：河南工程学院计算机学院;河南图像识别工程技术中心;
【基金】：国家自然科学基金(61301232)~~
【分类号】：TP391.41

【相似文献】