摘要
针对现有方法在哈希函数构造过程中没有考虑数据的稀疏结构,提出了一种基于稀疏重构的哈希函数学习方法。利用相似点的l_(21)范数对重构系数进行了稀疏约束,以增强局部保持映射过程中的判别性,并构建拉普拉斯矩阵进行局部邻域关系的约束,在调和协方差矩阵和最小化数据的重构误差间建立了一种平衡机制。从特征所在的空间与经映射后的汉明空间的可判别性角度出发,对哈希函数构造过程中的内在要求和约束同时进行了考虑并综合权衡。采用公共图像检索数据集Caltech-256进行实验,实验结果表明:32位编码长度时,本文算法的检索精度比其他无监督的深度哈希算法至少提高了4.69%。
Because the sparse structure of data was not considered in the construction of hash function by the existing methods,a hash function learning method based on sparse reconstruction was proposed. l_(21) norm of the similar point was used to sparsely restrain reconstruct coefficient to enhance the discriminat property in the process of locality preserving mapping. The Laplace matrix was used to constrain the local neighborhood relation. A balance mechanism between the covariance matrix and the reconstruction error of minimize data was established. The inherent requirements and constraints in the process of hash function construction were considered and integrated from the perspective of the feature space and the mapped Hamming space. A public image retrieval data set Caltech-256 was used to test. The experimental results show that when the encoding length is 32 bit,the retrieval accuracy of the proposed algorithm is 4. 69% higher than that of other unsupervised deep hashing algorithm.
出处
《河南科技大学学报（自然科学版）》
CAS
北大核心
2018年第3期77-82,共6页
Journal of Henan University of Science And Technology:Natural Science
基金
国家自然科学青年基金项目(61303028)
关键词
哈希函数
图像检索
稀疏重构
稀疏检索
汉明空间
hash function
image retrieval
sparse reconstruction
sparse retrieval
hamming space
