基于离散图哈希的图像检索算法研究
发布时间：2018-08-29 16:52
【摘要】：随着互联网的高速发展,图像数据规模呈爆发式增长,研究如何在海量图像数据中高效检索出用户感兴趣的图像具有重大意义。传统的基于内容的图像检索算法主要采用图像的内容特征进行相似度匹配,存在特征之间“语义鸿沟”、特征维度高、存储空间大、检索效率低等问题,用基于哈希的图像检索算法可以有效弥补上述不足。目前基于哈希的图像检索算法检索准确率和检索效率还无法令人满意,本课题为解决这两个问题,提出基于离散图哈希的图像检索算法,主要研究成果如下:针对图像检索中的特征提取问题,提出一种基于拉普拉斯图模型的多视图非负特征融合算法。采用非负矩阵分解技术,将每一种视图特征进行特征转换,分解后的非负特征表达能力更强。每种视图特征仅仅是图像语义信息的一个表现层面,采用拉普拉斯图模型将多种视图的非负特征嵌入到一个统一的潜在空间,在该空间中融合后的特征可以更完整地表达图像语义特征。在构造拉普拉斯图模型时,引入“锚点”技术,降低拉普拉斯矩阵的计算复杂度。根据算法模型,在公开数据集上进行了实验,验证本文研究的多特征融合算法要比单特征检索准确率高。为了构建高效索引,提出了一种有监督机器学习的离散图哈希图像检索算法。通过学习哈希函数,将原始空间中的数据特征映射到汉明空间,保持数据相似性,在汉明空间中计算哈希码相似度。在学习哈希函数时,利用数据的标签信息对图像语义信息的表示作用,采用有监督的机器学习方法,使用离散图优化框架,直接对约束变量为离散值的目标函数优化,避免了传统方法采用“松弛”策略导致哈希编码质量较低,提高了检索精度,采用离散循环坐标下降法,逐位生成所有训练样本的哈希码,提高了哈希码的生成速度,使用该算法在公开数据集上与其他主流哈希方法进行对比实验,验证了本文提出的哈希算法在图像检索时的高效性。为了验证多视图非负特征融合后的特征有效性,用离散图哈希算法对该特征进行检索实验,实验表明该特征与离散图哈希算法结合,在图像检索中能够提高检索准确性。
[Abstract]:With the rapid development of the Internet, the scale of image data is increasing explosively. It is of great significance to study how to efficiently retrieve the images of interest to users in the massive image data. The traditional content-based image retrieval algorithm mainly uses the content feature of the image to carry on the similarity matching, which has the problem of "semantic gap" between the features, the feature dimension is high, the storage space is large, the retrieval efficiency is low, and so on. Hash-based image retrieval algorithm can effectively compensate for the above deficiencies. At present, the accuracy and efficiency of hash based image retrieval algorithm are not satisfactory. In order to solve these two problems, this paper proposes an image retrieval algorithm based on discrete image hashing. The main research results are as follows: aiming at the feature extraction problem in image retrieval, a multi-view non-negative feature fusion algorithm based on Laplace diagram model is proposed. The non-negative matrix decomposition technique is used to transform the features of each view, and the non-negative feature expression is stronger after the decomposition. Each view feature is only a representation of the semantic information of the image. The Laplace diagram model is used to embed the non-negative features of multiple views into a unified potential space. The fused features in this space can express image semantic features more completely. In order to reduce the computational complexity of Laplace matrix, the "anchor point" technique is introduced in the construction of Laplace diagram model. According to the algorithm model, experiments are carried out on the open data set to verify that the multi-feature fusion algorithm studied in this paper is more accurate than the single-feature retrieval algorithm. In order to construct an efficient index, a supervised machine learning discrete image retrieval algorithm is proposed. By learning the hash function, the data feature in the original space is mapped to the hamming space, and the similarity of the data is maintained, and the similarity of the hash code is calculated in the hamming space. When learning hash function, we use label information of data to represent semantic information of image, adopt supervised machine learning method, use discrete graph optimization framework, and directly optimize objective function with constraint variable as discrete value. The traditional method of "relaxation" strategy is used to avoid the lower hash coding quality and improve the retrieval accuracy. The discrete cyclic coordinate descent method is used to generate the hash codes of all training samples bit by bit, and the generation speed of the hash codes is improved. The proposed algorithm is compared with other popular hash methods on the open data set, and the efficiency of the proposed hash algorithm in image retrieval is verified. In order to verify the feature validity of multi-view non-negative feature fusion, a discrete Tuhash algorithm is used to retrieve the feature. The experiment results show that the feature can improve the retrieval accuracy in image retrieval by combining the feature with the discrete image hash algorithm.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
本文编号：2211817
[Abstract]:With the rapid development of the Internet, the scale of image data is increasing explosively. It is of great significance to study how to efficiently retrieve the images of interest to users in the massive image data. The traditional content-based image retrieval algorithm mainly uses the content feature of the image to carry on the similarity matching, which has the problem of "semantic gap" between the features, the feature dimension is high, the storage space is large, the retrieval efficiency is low, and so on. Hash-based image retrieval algorithm can effectively compensate for the above deficiencies. At present, the accuracy and efficiency of hash based image retrieval algorithm are not satisfactory. In order to solve these two problems, this paper proposes an image retrieval algorithm based on discrete image hashing. The main research results are as follows: aiming at the feature extraction problem in image retrieval, a multi-view non-negative feature fusion algorithm based on Laplace diagram model is proposed. The non-negative matrix decomposition technique is used to transform the features of each view, and the non-negative feature expression is stronger after the decomposition. Each view feature is only a representation of the semantic information of the image. The Laplace diagram model is used to embed the non-negative features of multiple views into a unified potential space. The fused features in this space can express image semantic features more completely. In order to reduce the computational complexity of Laplace matrix, the "anchor point" technique is introduced in the construction of Laplace diagram model. According to the algorithm model, experiments are carried out on the open data set to verify that the multi-feature fusion algorithm studied in this paper is more accurate than the single-feature retrieval algorithm. In order to construct an efficient index, a supervised machine learning discrete image retrieval algorithm is proposed. By learning the hash function, the data feature in the original space is mapped to the hamming space, and the similarity of the data is maintained, and the similarity of the hash code is calculated in the hamming space. When learning hash function, we use label information of data to represent semantic information of image, adopt supervised machine learning method, use discrete graph optimization framework, and directly optimize objective function with constraint variable as discrete value. The traditional method of "relaxation" strategy is used to avoid the lower hash coding quality and improve the retrieval accuracy. The discrete cyclic coordinate descent method is used to generate the hash codes of all training samples bit by bit, and the generation speed of the hash codes is improved. The proposed algorithm is compared with other popular hash methods on the open data set, and the efficiency of the proposed hash algorithm in image retrieval is verified. In order to verify the feature validity of multi-view non-negative feature fusion, a discrete Tuhash algorithm is used to retrieve the feature. The experiment results show that the feature can improve the retrieval accuracy in image retrieval by combining the feature with the discrete image hash algorithm.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41
【参考文献】
相关期刊论文 前1条
1 欧新宇;伍嘉;朱恒;李佶;;基于深度自学习的图像哈希检索方法[J];计算机工程与科学;2015年12期
相关博士学位论文 前1条
1 金仲明;基于哈希算法的海量多媒体数据检索研究[D];浙江大学;2015年
，本文编号：2211817
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2211817.html
