非负矩阵分解耦合视觉词典的图像检索算法  

Image Retrieval Algorithm Based on Non-negative Matrix Factorization Coupled Visual Dictionary

作　　者：李峰[1] 应帅 卢文超 LI Feng;YING Shuai;LU Wen-chao(College for Creative Studies, Changzhou Vocational Institute of Textile and Garment, Changzhou 213164, China;School of mechanical and Electronic Engineering, Ji＇an College, Ji＇an 343000, China)

机构地区：[1]常州纺织服装职业技术学院创意学院,常州213164 [2]吉安职业技术学院机械与电子工程学院,吉安343000

出　　处：《包装工程》2018年第17期215-222,共8页Packaging Engineering

基　　金：江苏省自然科学基金(BK20151191)

摘　　要：目的解决当前图像检索技术中,图像特征稀疏编码收敛速度慢,以及局部特征空间信息不足易导致检索误差较大等问题,提出一种基于l0稀疏约束非负矩阵分解耦合视觉词典优化的图像检索算法。方法首先,在非负矩阵分解(Non-negative Matrix Factorization,NMF)的基础上,对系数矩阵设置l0个约束来限制其稀疏性,从而定义一种l0稀疏约束的NMF方法。再通过一种自适应序列词典初始化方案,从训练样本获得词典的初始估计。然后,利用l0稀疏约束的NMF来增强视觉词典,对图像局部描述符进行稀疏编码,并利用最大池化操作来生成聚合特征向量,从而保留局部描述符的关键属性。最后根据得到的特征向量,引入Minkowski距离来衡量查询图像与数据库的相似性,输出检索图像。结果实验结果表明,与当前图像检索方案相比,所提算法具有更高的查准-查全率和收敛速度。结论所提算法返回的图像与查询图像相似度高,在包装商标检索等领域具有一定的参考价值。The work aims to solve such defects as the slow convergence speed of image feature sparse coding, and large retrieval error induced by insufficient local feature space information in current image retrieval technology. An image retrieval algorithm based on l0 sparse constraint non-negative matrix factorization coupled visual dictionary optimization was proposed. Firstly, on the basis of the non-negative matrix factorization（NMF）, l0-constraints were stetted on the coefficient matrix to limit its sparsity, so that a NMF framework for l0-sparse constraints was defined. Then, an initialization scheme of adaptive sequence dictionary was proposed to obtain the initial estimation of the dictionary from the training samples. Then, the NMF with l0 sparse constraints was used to enhance the visual dictionary for sparse coding on image local descriptors, and the polymerization feature vectors were generated by the maximum pooling operation to retain the key attributes of the local descriptor. Finally, according to the obtained feature vectors, the Minkowski distance was introduced to measure the similarity between the query image and the database for outputting the retrieval image. The experimental results showed that the proposed algorithm had a higher precision-recall rate and faster convergence speed compared with the current image retrieval scheme. The image retrieved by the proposed algorithm has a high similarity to the query image, which has a certain reference value in the fields of package trademark retrieval, etc.

关 键 词：图像检索 非负矩阵分解 视觉词典 稀疏编码 最大池化 Minkowski距离 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]