Abstract: There is no clear space structure among images, so correlation information of images' space structure could not be utilized effectively. To tackle this problem, a new space relationship feature based image retrieval method was proposed. Firstly, feature vectors were extracted from all images which contain the queried image. And then, all of the every two feature vectors' similarities were calculated to form a similarity matrix. The set of similarity matrix's columns was taken as the new feature vector, namely the new space relationship feature vector, so the former feature vectors could be mapped into a Euclidean space. Finally, similarities were calculated on the new feature space. In this way, the problem of feature vectors' similarity was turned into the new space relationship feature vectors' similarity. The space structure among images was clearer than before on the new feature space, so the accuracy of image retrieval was improved. The experimental results on the Corel database show that the average retrieval precision, recall-precision and visual evaluation metric of the proposed method have advantage over the color histogram in the image retrieval task. The proposed image retrieval method based on the new space relationship feature sufficiently utilizes correlation information of images' space structure and it has better retrieval effect.

Key words: space structure, feature vector, similarity matrix, feature space, image retrieval

摘要： 图像与图像之间没有清晰的空间结构，这样就不能有效利用图像间空间结构上的相关性信息，针对此问题提出一种基于新的空间关系特征的图像检索方法。首先，提取待查询图像在内的全部图像的特征向量。然后，计算特征向量每两个之间的相似性，形成相似性矩阵。将相似性矩阵的列集合作为新特征向量，命名为新的空间关系特征向量，从而将原来的特征向量映射到一个欧氏空间上。最后，在新特征空间上计算相似性，特征向量之间的相似性问题就转化为新的空间关系特征向量之间的相似性问题。在新特征空间上，图像与图像之间的空间结构变得清晰了，有利于图像检索准确度的提高。在Corel数据库上进行实验，所提方法在平均检索查准率、查全率-查准率和可视化评价指标上都优于基于颜色直方图的图像检索方法。结果表明，基于新的空间关系特征的图像检索方法有效利用了图像间空间结构上的相关性信息，具有更好的检索效果。

关键词: 空间结构, 特征向量, 相似性矩阵, 特征空间, 图像检索