基于改进Fisher准则与极限学习机集成的图像隐写分析

An Image Steganalysis Algorithm Based on Improved Fisher Criterion and Extreme Learning Machine Ensemble

集成分类器是目前图像隐写分析中广泛使用的分类器。针对集成分类器中基分类器受离群样本影响较大,集成策略效果不佳的缺点,提出一种基于改进Fisher准则与极限学习机集成的图像隐写分析算法。首先,通过重新定义类内散度矩阵以提高Fisher准则模型的准确性,之后基于改进的Fisher准则并使用Bagging算法训练若干基分类器,最后使用极限学习机作为元分类器来建立基分类器集合与正确决策之间的联系。实验结果表明,在不同的隐写算法与嵌入率的条件下,与传统集成分类器和基于选择性集成的集成分类器相比,所提算法降低了3.5%与1.8%的检测错误率,说明能够有效提高集成分类器的检测精度。

The ensemble classifiers is one of the most widely used image steganalysis methods now.Since the base learners of ensemble classifiers were influenced by outlier samples highly and the effect of fusion method was bad, an image steganalysis algorithm based on improved Fisher criterion and extreme learning machine ensemble was proposed.First, the new method improved the accuracy of Fisher model through redefining the within class scatter, then some base learners were generated based on improved Fisher criterion and Bagging algorithm, at last, an extreme learning machine was trained as a metalevel classifier to learn the relationship between base learners decisions and true decisions.Experimental results show that the average error rate of the proposed method decreases by 3.5% and 1.8% in comparison to typical ensemble classifiers and ensemble classifiers based on selective ensemble, therefor demonstrating the proposed method could improve the accuracy of ensemble classifiers.