基于多尺度特征变换与颜色相关性的商标检索算法  

Trademark Retrieval Algorithms Based on SIFT and Color Correlation

作　　者：钟瑞泽 ZHONG Rui-ze(School of Arts and Humanities,Guangzhou Academy of Fine Arts,Guangzhou 510260,China)

机构地区：[1]广州美术学院艺术与人文学院,广州510260

出　　处：《包装工程》2018年第23期200-208,共9页Packaging Engineering

基　　金：广东省"十二五"规划科学研究项目(2012JK190)

摘　　要：目的提出一种快速有效的商标注册相似性检查方法,以解决当前基于SIFT的商标检索系统易出现漏检、误检,导致检索精度不高的问题。方法首先,利用SIFT进行尺度空间创建,并检测商标的特征关键点,通过确定关键点的主方向,可得到具有旋转、缩放、平移、视图变化不变性的图像形状特征描述符。随后,根据像素与其邻域的颜色和空间位置,定义一种改进的颜色相关性,为了有效避免不同商标可能具有相似的颜色特征,对不同的颜色赋予一个权重因子,从而得到一个反映颜色空间相关性与颜色排布疏密度的颜色特征。然后,将SIFT与颜色相关特征向量进行加权组合,并根据实际过程中占主导作用的特征来改变权重。最后,根据加权组合特征,引入马氏距离对查询商标与数据库商标进行相似度量,输出检索商标。结果实验结果表明,与当前先进的商标检索系统对比,所提算法具有更高的检索准确性与效率。结论所提算法具有良好的检索准确率与鲁棒性,在商标注册等领域具有一定的实用价值。ABSTRACT： The work aims to propose a rapid and effective trademark registration similarity inspection method to solve the problem of low retrieval accuracy induced by missing detection and wrong detection in current SIFT-based trademark retrieval system. Firstly, SIFT was used to create the scale space and detect the key points of trademark feature. Then, the main direction of the key points was determined to obtain the image shape characteristic descriptors with rotation, scaling, translation and view change invariance. Subsequently, according to the color and spatial location of pixels and their neighbors, an improved color correlation was defined to effectively avoid the possibility that different trademarks might have similar color features. A weight factor was applied to different colors, and thus a color feature reflecting the color space correlation and color arrangement density was obtained. Then, SIFT and color correlation feature vectors were weighted and combined, and the weights were changed according to the dominant features in the actual process. Finally, according to the weighted combination features, Mahalanobis distance was introduced to measure the similarity between query trademarks and database trademarks, and the retrieved trademarks were outputted. The experimental results showed that, compared with the current advanced trademark retrieval system, the proposed algorithm had higher retrieval accuracy and efficiency. The proposed algorithm has good retrieval accuracy and robustness which has certain practical value in trademark registration and other fields.

关 键 词：商标检索 多尺度特征变换 颜色相关性 特征关键点 主方向 马氏距离 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]