互信息启发的相似度组合图像检索算法

Combining similarity measures in content-based image retrieval guided by mutual information

图像的视觉特征与用户描述之间的差距一直是影响基于内容的图像检索准确度的最主要因素。对多种相似度进行组合来检索图像是近几年图像检索领域涌现出的一个研究热点,也是缩小这种差距的一种有效途径。如何选择更好的组合方法则是该领域很多研究者关注的核心问题。提出一种新的相似度组合算法。该算法基于互信息度量相对熵的原理,计算连续变量相似度与离散变量相似性之间的相关性,对多种相似度进行选择,以"和规则"组合相似度。在公用数据集上进行检索实验,该算法优于当前其他的"和规则"下的组合方法。

The lack of accordance between the information that one can extract from an image and the interpretability of the same image in a given situation is the most important factor that hampers the accuracy of content-based image retrieval （CBIR）. Recently, the combination of several similarity measures draws much interest in the CBIR area, It can be shown that is effective in reducing this discordance. The core problem is：how to choose a better way to combine these similarities？ In this paper, we propose a new combination algorithm. It combines similarity measures under the sum rule based on mutual information which estimates the correlation between the continuous random variable similarity measures and the discrete random variable similarity. The experimental results show that this algorithm achieves a high accuracy and efficiency in real-world image collections.