综合语义特征和视觉特征的二阶段纹理图像检索

Two-Stage Texture Image Retrieval by a Combination of Semantic Feature and Visual Feature

提出一种融合模糊语义概念和精细视觉特征的纹理图像检索方法.首先根据语言表达式和模糊语义概念对整个图像库进行快速有效的粗搜索,得到具有"软"边界的语义检索结果;然后根据视觉特征在语义检索结果中(而不是整个图像库)进行精细的检索.该方法很好地结合了基于内容的图像检索和基于语义的图像检索两者的优势,使得用户既可以根据语义概念对图像库进行快速浏览和检索,也能根据查询用例图像的视觉特征进行精细的匹配;另外,由粗到细的二阶段策略也明显地提高了其检索性能.在Brodatz纹理库中的实验结果表明,通过调整合适的语义检索边界,该方法的检索性能明显优于基于视觉特征的图像检索方法.

By integrating the fuzzy semantic concept and fine visual features, we put forward a two-stage image retrieval （TSIR） methodology. In the first stage, TSIR quickly searches the whole image base according to fuzzy semantic feature, and obtains a semantic search result that characterizes with a ＂soft＂ boundary. In the second stage, TSIR retrieves the more accurate images from the semantic search result rather than from the whole collection according to the visual feature. TSIR possesses both the advantages of content-based image retrieval and semantic-based image retrieval. Users of TSIR can quickly browse and search the whole image base with semantic concept. And moreover, they can finely retrieve the more accurate images according to the visual feature of query example. Furthermore, being characterized with the coarse-to-fine strategy, TSIR achieves the improvement of retrieval performance. The experiments based on Brodatz image set show that the retrieval performance of TSIR outperforms the traditional content-based image retrieval system.