基于区域生长耦合多分类器的商标检索算法

目的针对商标检索算法中易出现的语义鸿沟,底层视觉特征与高层语义相关性不强而导致商标检索精度不理想的问题,定义一种基于区域生长耦合多分类器的商标检索方案。方法首先对输入的商标进行预处理,去除图像中的噪声和杂散点,并通过3D直方图和聚类算法来提取输入图像中的主颜色;基于区域生长算法,合并具有相同颜色标签的所有连接点,以形成颜色区域;然后根据生成的颜色区域,分别定义颜色分类器、形状分类器和关系分类器,利用每个分类器计算查询图像和数据库中图像的检索优势概率;最后通过决策组合,根据检索规则和列表长度找到最相似的商标,并利用动态选择方案进一步提高检索准确率。结果实验结果表明,与当前商标检索方案相比,所提检索系统具有更为理想的Precision-Recall曲线,对缩放、扭曲和噪声具有更高的鲁棒性。结论所提方案在各类几何变换下具备较高的检索准确率,对商标注册、版权保护等行业有较好的借鉴意义。

Trademark Retrieval Based on Region Growth Coupled Multi-classifier

The work aims to define a trademark retrieval scheme based on region growth coupling multi-classifier for the semantic gap in the trademark retrieval algorithm and the low accuracy of trademark retrieval due to low correlation between the underlying visual features and the high-level semantics leads to. Firstly, the input trademark was preprocessed to remove noise and spurious points in the image. The main color of the input image was extracted by 3D histogram and clustering algorithm, and the region growing algorithm was implemented to merge all the join points with the same color label to form the color region. Secondly, color classifier, shape classifier and relational classifier were defined based on the generated color region. Each classifier was used to calculate the retrieval advantage probability of the query image and the image in the database. Finally, through the decision-making combination process, the most similar trademarks were found according to the retrieval rules and the length of the list, and the dynamic selection scheme was used to further improve the system performance. Through the experiments, compared with current trademark retrieval schemes, the proposed retrieval system had more ideal Precision-Recall curve, which had higher robustness to scaling, distortion and noise. This algorithm has high retrieval accuracy under various geometric transformations, which has a good reference value for trademark registration, copyright protection and other industries.