基于多特征多核哈希学习的大规模图像检索

发布时间：2019-06-10 01:39

【摘要】：哈希学习方法解决了图像大数据的检索效率低,存储代价高的问题.目前已有的核哈希方法中,要么仅使用一种特征对应单个核函数,要么是多特征对应单个核函数,它们忽视了综合考虑不同的核函数具有的不同作用和不同的特征包含不同的信息的事实.本文提出了一种自适应的多特征多核的哈希学习算法(MFMKH),该算法能够自适应学习多特征融合的权重系数和多核融合的权重系数,将多特征和多核的优点进行了双重融合.本算法中的特征融合解决了单特征所包含的信息量单一不足的问题,采用多种不同的核函数能够弥补单核学习能力上的不足,具有多特征自适应融合和多核学习的双重优点.在标准的IRMA,Ultrasound和Cifar10数据集上的实验表明,本文算法检索性能明显优于同类基于核的哈希学习方法,且与监督的深度哈希相比训练时间显著少的情况下检索性能在Cifar10数据集上是可竞争的.
[Abstract]:The hash learning method solves the problems of low retrieval efficiency and high storage cost of image big data. In the existing kernel hash methods, either only one feature is used to correspond to a single kernel function, or a multi-feature corresponds to a single kernel function. They ignore the fact that different kernel functions have different functions and different characteristics contain different information. In this paper, an adaptive multi-feature and multi-core hash learning algorithm (MFMKH), is proposed, which can adaptively learn the weight coefficient of multi-feature fusion and the weight coefficient of multi-core fusion, and combine the advantages of multi-feature and multi-core. The feature fusion in this algorithm solves the problem of single information contained in single feature. Using many different kernel functions can make up for the deficiency of single kernel learning ability, and has the dual advantages of multi-feature adaptive fusion and multi-core learning. Experiments on standard IRMA,Ultrasound and Cifar10 datasets show that the retrieval performance of this algorithm is obviously better than that of similar kernel-based hash learning methods. Compared with the supervised deep hash, the retrieval performance is competitive on the Cifar10 dataset when the training time is significantly less.
【作者单位】：重庆邮电大学计算机科学与技术学院;重庆邮电大学计算智能重庆市重点实验室;
【基金】：国家重点研发计划云计算和大数据重点专项(批准号:2016YFB1000905) 国家自然科学基金项目(批准号:61672120) 重庆自然科学基金项目(批准号:cstc2015jcyj A40036)资助
【分类号】：TP391.41

【相似文献】