SVM及其鲁棒性研究

多数人脸识别方法是利用大量正确标记的训练样本来学习精度足够高的识别模型。收集人脸图像并对其进行正确的标记会耗费大量的人力、物力,为了给已有的图像进行标注,研究者进行了大量的工作,但由于多种原因,标记的图像不一定全部正确,称这种标记错误为类别噪声。文中针对含类别噪声的人脸识别问题,指出SVM适用于这类问题,并通过分析位于不同位置的样本对分类的影响从理论上解释了SVM对噪声具有鲁棒性的原因。在SVM基础上,删除一定比例的被判定为噪声的样本后,鲁棒性能有所提高。PubFig数据集上的量实验验证了SVM及改进算法在含类别噪声学习中的有效性。

Research on the Robust SVM

Most methods of face recognition use amounts of corrected labeled samples to learn recognition models with high curate.Collecting face images and labeling them will consume plenty of manpower.In order to label the possession images,researchers have done many works and have made many contributions,but due to personal reason,the labels may be not correct entirely,we call the incorrect labels class noise.The paper is aimed at face recognition with class noise,point that SVM suit for these problems and explain the reason why SVM is robust to noise according to influence of support vectors’ position to classification.Discarding certain samples which were judged as noise based on SVM improves the robustness.Amounts of experiences in PubFig dataset verify the efficiency of SVM and the improvement algorithm in noisy learning.