论文部分内容阅读
人脸识别是人工智能领域研究的重要方向.传统人脸识别方法大多考虑降低识别错误率,忽略了不同误分类代价的差异性.此外,训练一个精确的人脸识别分类器需要大量的标记样本,实际问题中标记样本往往很少.为解决这一问题,文章提出了一种基于高斯混合模型和EM算法的增量序贯三支决策方法,用于解决标记样本不足的代价敏感的人脸识别问题.文章针对人脸图像数据建立高斯混合模型,通过EM算法将有标记和未标记的图像数据进行综合识别,充分利用未标记样本可以帮助训练的分类器更好地表示测试样本,并得到更理想的识别效果.设计动态代价敏感识别策略,不仅考虑降低识别错误率,而且实现识别错误率和误分类代价之间的平衡.当标记样本不足时,引入边界域决策.随着序贯分类决策过程中可用信息的增加,边界域决策可以转化为正域决策或负域决策,从而形成增量式序贯三支决策过程,并通过实验验证了其有效性.“,”The accuracy of face recognition has been able to reach a very high level in recent years.Howev-er,many studies on face recognition usually focus on minimizing the number of recognition mistakes while neglecting the imbalance of different misclassification costs,which is not really reasonable in many real-world scenarios.In addition,training a precise classifier often requires a large number of labeled samples,and there is only few labeled samples in many cases.To address this problem,we propose an incremental sequential three-way decision method for cost-sensitive face recognition with insufficient labeled samples,in which a Gaussian mixture model and expectation maximization(EM)algorithm are used to label new samples.We construct a Gaussian mixture model for the whole data of the face images,which mixed the labeled and unlabeled image data via EM algorithm.In many practical applications,the labeled samples are scarce and insufficient.So taking full advantage of the unlabeled samples can help the learned classifier well represent the testing samples and give better results.Instead of minimizing the recognition error rate only,we seek for a balance between recognition error rate and misclassification costs.And we consider a delayed decision,also called the boundary decision,when the labeled images are insufficient.With the in-creasing available information in the sequential process,the boundary decision can be converted to positive decision or negative decision,thus forming an incremental sequential three-way decision process.The use of sequential three-way decision can decrease the decision costs effectively.The experiments demonstrate the effectiveness of this method.