基于模糊关联规则和决策树的图像自动标注

Automatic image annotation based on fuzzy association rules and decision trees

传统的基于关联规则算法的图像自动标注存在"锐利边界"问题,使分类存在模糊性、不准确性。且随着多媒体技术的飞速发展,图像信息数据迅速增长,海量的图像数据会形成大量冗余的关联规则,这将导致分类效率大大降低。针对这2个问题,文中提出基于模糊关联规则和决策树的图像自动标注模型。该模型首先获得关联训练图像低层特征和高层语义的模糊关联规则,再利用决策树方法删减冗余的模糊关联规则,基于决策树删减后的模糊关联规则,大大减小了算法的计算复杂度。实验在Corel 5k和IAPR-TC12两个基准数据集上进行,并从精度、召回率、F-measure以及产生的规则数量几个度量措施上进行比较。与其他几种前沿的图像自动标注方法的结果对比表明,该方法在图像的标注精度和标注效率上有很大的提高。

The traditional automatic image annotation based on association rules exists the problem of sharp boundary,which makes classification more fuzzy and inaccurate. Moreover,with the rapid development of multimedia technology,the size of image data increases quickly. Massive image data will produce a lot of redundant association rules,which greatly decreases the efficiency of image classification. In order to solve these two problems,this paper proposes an automatic image annotation approach based on fuzzy association rules and decision trees. The approach firstly obtains fuzzy association rules which represent the fuzzy correlations between low-level visual features and high-level semantic concepts of training images. Then,decision tree is adopted to reduce the redundant fuzzy association rules. As a result,computational complexity of the algorithm is decreased to a large degree. Experiments were done on Corel5 k and IAPR-TC12 datasets. The evaluation measures are compared from the aspects of precision,recall,F-measure and the number of rules. The experimental results show that the proposed method acquires higher accuracy and efficiency in comparison with several state-of-the-art automatic image annotation approaches.