用于人脸情感识别的低冗余稀疏性集成剪枝

Sparsity and low-redundancy ensemble pruning for facial expression recognition

为降低噪声和异常值对集成剪枝的影响,并鲁棒性地选择更稀疏的基分类器子集,从而提升人脸情感识别的性能,提出了一种具有依赖分数的鲁棒性稀疏低冗余集成剪枝方法用于人脸情感识别。首先,该方法将样本实例的预测结果视为基分类器特征,分别利用互信息和熵来评估成对基分类器之间的依赖性和它们之间的优先级。其次,将优先级依赖关系添加到基于回归的目标方程中实现冗余基分类器的修剪,此目标方程使用l_(2,1)范数来增加分类器子集的鲁棒性从而提升算法的泛化性能。然后,将内积正则化项引入到目标方程中,通过计算分类器特征系数向量内积的绝对值的和去选择稀疏和低冗余的基分类器。最后,使用大多数投票法对选择的基分类器子集进行集成从而得到最终的识别结果。结果表明:本文提出的方法在FER2013、JAFFE、CK+和KDEF 4个公共人脸情感数据集上的识别准确率,比所有基分类器进行集成得到的准确率分别高3.29%、10.39%、1.76%和4.89%,表明该方法可以选择出识别效果更好、冗余度更低的分类器子集,提高集成剪枝的泛化能力。

To reduce the impact of noise and outliers on ensemble pruning and robustly select a more sparse subset of base classifiers to improve the performance of facial expression recognition,a robust sparse and low-redundancy ensemble pruning method with dependency scores is proposed in this paper for facial expression recognition.First,the method treats the prediction results of sample instances as base classifier features,and evaluates the dependency and priority between pairs of base classifiers using mutual information and entropy,respectively.Second,the priority dependency is added to the regression-based objective equation to prune redundant base classifier,which uses the l_(2,1)-norm to increase the robustness of the classifier subset and thus improve the generalization performance of the algorithm.Then,an inner product regularization term is introduced into the target equation to select sparse and low-redundant base classifiers by computing the sum of the absolute values of the inner products of the classifier feature coefficient vectors.Finally,the majority voting method is used to integrate the selected subset of base classifiers to obtain the final recognition results.The recognition accuracy of the proposed method on four public facial expression datasets,FER2013,JAFFE,CK+,and KDEF,is 3.29%,10.39%,1.76%,and 4.89%higher than those obtained by integrating all base classifiers,respectively,indicating that the method can select a subset of classifiers with better recognition results and lower redundancy,and improve the ensemble pruning generalization ability.