基于散布矩阵分析的相关反馈算法及应用

Relevance feedback approach based on scatter matrix analysis

相关反馈技术是一种较常用的提高信息检索精度的方法。在图像检索领域,相关反馈技术被认为是解决图像高层语义内容和低层视觉特征之间差异的一种有效方法。视觉特征的权值调整是一类应用较多的相关反馈技术,权值调整方法中存在矩阵奇异问题,本文提出了一种新的基于散布矩阵分析的相关反馈算法,解决了矩阵奇异问题。该方法通过分析与检索目标相关图像在特征空间中的散布来构造目标图像类的投影空间,该空间对应于一个高层语义类在特征空间中分布密集的子空间,在投影空间中计算相似图像;同时根据每次反馈的信息不断修正投影空间来提高系统的检索性能。在Corel图像数据库中的实验结果表明该算法具有良好的检索性能。

Relevance Feedback is a technique to improve information retrieval performance, and an effective approach to bridge the gap between high-level semantic concepts and low-level visual features in content-based image retrieval. Although re-weighting of image visual features are frequently used, the matrix singularity problem of the re-weighting approach is still unresolved. A novel relevance feedback approach based on scatter matrix analysis to the matrix singularity problem is proposed. By analyzing the dispersion of the images relevant to retrieval goal in the feature space, similar images are calculated in a projected space, corresponding to this projected space of the feature space where the images belonging to one semantic group distribute more closely. Moreover, the projected space is adjusted to feedback each round, which improves the system＇s retrieval performance. Experimental results on Corel image collection show that this algorithm achieves a better retrieval performance.