基于卷积神经网络和LSH的图像检索算法  

Image retrieval algorithm based on convolutional neural network and Hash algorithm

作　　者：杨荣 张建刚[2] 贾晖 YANG Rong;ZHANG Jiangang;JIA Hui(School of Computer Science and Technology,Xi'an University of Posts and Telecommunications,Xi'an 710121,China;Power Station Information and Monitoring Technology Department,Xi'an Thermal Power Research Institute Co.,LTD,Xi'an 710032,China)

机构地区：[1]西安邮电大学计算机学院,陕西西安710121 [2]西安热工研究院有限公司电站信息及监控技术部,陕西西安710032

出　　处：《西安邮电大学学报》2022年第2期88-94,共7页Journal of Xi’an University of Posts and Telecommunications

基　　金：陕西省教育厅科学研究计划项目(103/205040120)。

摘　　要：为了提高图像检索的准确度和检索效率,提出一种基于卷积神经网络和局部敏感哈希(Locality-Sensitive Hashing,LSH)算法的图像检索算法。使用图像库ImageNet对视觉几何小组16(Visual Geometry Group 16,VGG16)网络进行训练,获取初始化参数。以卷积神经网络为基础,增加哈希层代替VGG16全连接层,获取图像的高维特征向量。利用哈希函数满足p-稳定分布的LSH算法将高维特征向量映射为哈希码,并将相似图像映射到同一个哈希桶中作为粗检候选集,计算并排序候选集中特征向量欧氏距离完成图像检索,从而得到最终的检索结果。实验结果表明,与其他基于不同哈希算法的图像检索算法相比,所提算法具有较高的准确性和较快的检索速度。In order to improve the accuracy and efficiency of image retrieval,an image retrieval algorithm based on the convolutional neural network and locality-sensitive Hashing algorithm is proposed.Use image library ImageNet to train visual geometry group 16(VGG16)network and obtain initialization parameters.Based on the convolutional neural network,the Hash layer is added to replace the VGG16 full connection layer,and the high-dimensional feature vectors of the images are obtained.The high-dimensional feature vectors are mapped into Hash codes by using the LSH algorithm whose Hash function satisfies the p-stable distribution,the similar images are mapped to the same Hash bucket as a candidate set for rough detection.The Euclidean distance of feature vectors in the candidate set is calculated and sorted to achieve image retrieval and the final retrieval result is obtained.Compared with other image retrieval algorithms based on different Hash algorithms,the experimental results show that the proposed method has higher accuracy and faster retrieval speed.

关 键 词：图像检索 卷积神经网络 局部敏感哈希算法 高维特征向量 

分 类 号：TP311.52[自动化与计算机技术—计算机软件与理论]