Abstract: Based on the feature of trademark image, a retrieval algorithm for trademark is proposed which combines the global feature with the local feature of images. The global feature reflects the overall information of the image that can help to build the candidate image database quickly, while the local feature can be matched with the candidate images more accurately. Extract HU invariant moments of the retrieved image and sort them according to similarity. Based on this result, match the candidate images accurately through extracting the SIFT features. Experimental results show that this method not only keeps the well descriptive ability of SIFT features, but also reduces the complexity and the counting times that are required by fine matching.

Key words: Content Based Image Retrieval（CBIR）, trademark, HU invariant moments, Scale Invariant Feature Transform（SIFT）

摘要： 利用商标图像的形状特征，提出了一种融合图像全局特征和局部特征的商标检索算法。其中全局特征反应了图像的整体信息，这些信息可用来较快地建立候选图像库，而局部特征则可以更准确地与候选图像进行匹配。提取图像的HU不变矩进行初步检索，按相似度排序，在此结果集的基础上对候选图像通过提取SIFT特征进行精确匹配。实验结果表明，该方法既保持了SIFT特征的良好描述能力，又减少了精确匹配需要的计算次数，降低了复杂度。

关键词: 基于内容的图像检索, 商标, HU不变矩, 尺度不变特征变换（SIFT）