传统的基于内容的目标检索方法一般通过提取和比较局部特征,往往计算复杂度较高而且需要进行离线训练。因此,以最新的主方向模板(DOT)特征为基础提出新的实时目标检索方法。同时,利用搜索窗口中网格得分的空间分布情况构造相似度映射图,并采用多层次金字塔评分去除网格得分较为分散的错误窗口。利用以上改进实现的目标检索方法,真正能够无须离线训练即可从输入视频直接得到检索结果。实验结果表明,本文方法在提升检索性能的同时仍然保持主方向模板匹配的实时处理能力。

Traditional methods for Content-based Object Retrieval usually employ extracting and comparing local features, which are too computational demanding and require offline training process. Therefore, a novel real-time CBOR method based on dominant orientation templates (DOT) is proposed. Besides, the proposed method makes use of the spatial distributions of grid scores to construct likelihood maps and utilizes a pyramid-scoring strategy to dispose the false windows in which the grid scores are dispersed. With the above improvement, the implementation for CBOR is able to retrieve the query target directly from the videos without offline training. The results show the proposed method improves the retrieval performance while retaining real-time processing ability of DOT matching.