基于相关反馈的图像标注与检索算法

Image annotation and retrieval algorithm based on relevance feedback

为了提高图像标注与检索的性能,提出了一种基于区域分割与相关反馈的图像标注与检索算法。该算法利用视觉特征与标注信息的相关性,采用基于区域的视觉特征对每幅图像采用聚类方法获得其一组视觉相似图像。通过计算与其距离最近的前3个分类的相似度,然后对这些关键字概率向量进行整合,获得最适合该图像的关键字概率向量,对图像进行标注。利用用户的反馈信息,修正查询关键词与每个分类之间的关系,进一步提高图像检索的准确性。实验结果表明,提出的算法具有更高的查准率与查全率。

To improve the performance of image annotation and retrieval, an image annotation and retrieval algorithm was proposed based on region segmenting and relevance feedback. The proposed algorithm used the relevance of visual characteristics and annotation information, and obtained a group of similar photos by clustering based on visual characteristics of regions. Then it calculated the similarities between the region and the nearest three classes, and merged the keyword probability vector （KPV） to get the most appropriate KPV. The proposed algorithm also adopted the user＇s feedback information to adjust the weights between query words and each category to improve the accuracy of image retrieval. Simulation results prove that the proposed algorithm greatly improves precision and recall of image retrieval.