摘要
针对用于图像隐写分析的集成分类器集成规模较大,集成方法未体现基分类器之间预测性能的差异等问题,提出一种基于层次修剪和加权投票的图像隐写分析算法。利用Bagging(bootstrap aggregation)算法得到若干基分类器,使用基于差异度的层次修剪算法对基分类器进行选择性集成,剔除部分基分类器,提高分类器整体分类精度与整体差异度,根据Fisher距离计算各基分类器的投票权值,基于加权投票机制进行集成判决。实验结果表明,与传统完全集成的集成分类器相比,该算法减少了2/3的基分类器数量,降低了5%的检测错误率,在选择性集成的速度上,优于基于遗传算法的集成分类器。
Aiming at the problems of ensemble classifiers for steganalysis that the ensemble scale is large and the fusion method fails to reflect the different ability of base learners？ an image steganalysis algorithm based on hierarchical pruning and weighted vote was proposed. Some base learners were generated based on the Bagging algorithm and the hierarchical pruning algorithm based on diversity was used to delete some bad learners for improving the detection accuracy and diversity. Different weights of remaining base learners were given according to Fisher distance and the weighted vote method was used to get the results. Re-sults of experiments show that the algorithm can reduce two-thirds of numbers of base learners and decrease the error rate by 5% in the comparison experiments with ensemble classifiers based on traditional complete ensemble. And the selective ensemble speed is higher than ensemble classifiers based on genetic algorithm.
出处
《计算机工程与设计》
北大核心
2017年第9期2432-2437,共6页
Computer Engineering and Design
基金
国家自然科学基金项目(61379152)
关键词
隐写分析
集成分类器
层次修剪
加权投票
差异度
steganalysis
ensemble classifiers
hierarchical pruning
weighted vote
diversity
