基于内容的大数据量商标检索系统

Content-Based Retrieval System from Huge Trademark Databases

到目前为止,大数据量的图像检索依然是一个难题,提出了运行在一种大型数据库上的基于内容的快速的商标图像检索.首先,从商标图像中提取两种统计特征,然后采用概率主成分分析降维,生成特征字典—数据库中商标图像集的一个特征映射.在检索阶段,采用快速的层次检索来得到一个数目不定的候选集,再通过相关反馈进行不断的优化,将候选集的数目减少,直至符合检索要求.在国家商标局提供的30,0270商标图像上运行本系统,每一个查询时间不超过0.3秒.

Till now, many trademark retrieval systems have been proposed. Retrieval from huge databases is still a challenging problem. This paper presented a fast content-based retrieval system from huge trademark databases. First, we introduce two appropriate statistical features. In follow, probabilistic principal component analysis （PPCA） is used to reduce feature dimension. In query stage, a fast hierarchical retrieval scheme is taken to get a variable number of candidate set. The query results will be iteratively optimized through relevance feedback. In every iteratiwe process, the size of probability relevance set is reduced to a limited number. Experiment results on a database of 300,270 trademark images demonstrate that the proposed system is fast and efficient. A query process costs only O. 3 seconds