基于Fisher判别字典学习的可拒识模式分类模型  

Pattern Classification Model with Rejection Recognition Based on Fisher Discriminant Dictionary Learning

作　　者：廖重阳 张洋[1] 屈光中 毕云云[1] 

机构地区：[1]合肥工业大学计算机与信息学院,合肥230009

出　　处：《计算机工程》2016年第4期202-208,共7页Computer Engineering

基　　金：安徽省自然科学基金资助项目(128085MF91)

摘　　要：针对模式分类任务中测试样本存在未知类别输入的问题,在稀疏表示分类技术的基础上提出一种可拒识模式分类模型。该模型在字典学习的目标函数中加入Fisher判别约束,使样本在该字典下分解的系数具有较大的类间散度和较小的类内散度,将训练样本在已学习字典下进行分解,并把分解后的系数构建多个局部线性块,为已构建的线性块建立超球覆盖模型,用于描述训练类样本系数的分布状况。对于测试样本,根据在已学字典下的分解系数是否在训练样本系数的覆盖模型范围内,做出拒识或接受分类处理的判决。在MINST手写体数据库上的实验结果表明,该模型在保持较高正确识别率的同时,能对非训练类样本进行有效的拒识处理。Aiming at the problem that the test data do not belong to the training data in pattern classification task,under the sparse representation-based classification framework, this paper presents a pattern recognition with reject options model based on Fisher discriminant dictionary learning. This model adds the Fisher discriminant constraint to the objective function of dictionary learning, so that the sparse decomposition coefficients under the learned dictionary have a larger scatter between classes and a smaller scatter within one class, and builds several local linear manifold subspace for the coefficients of the training data to make an approximation of the nonlinear manifold space which the coefficients belong to. It makes a hypersphere cover to this built subspace to describe the distribution of the coefficients. For the test data, the model can make the decision accept the sample or reject it according to the hypersphere cover border. Experimental results on MINST show that the method has high correct recognition rate and can process with reject recognition to non training samples.

关 键 词：可拒识 字典学习 FISHER判别分析 基于稀疏表示的分类 流形 最大线性块 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]