最大似然稀疏编码在人脸识别中的研究

Research on Maximum Likelihood Sparse Coding in Face Recognition

稀疏编码(SRC)是一种用于人脸识别的方法,该方法把检测图像表示为一组训练样本的稀疏线性组合,表示的准确性通过L2或L1残余项来衡量。此模型假定编码残余项服从高斯分布或拉普拉斯分布,实际上却不能很准确地描述编码错误率。为了解决这个问题,提出了一种新的稀疏编码方法,建立一种有约束的回归问题模型,用最大似然稀疏编码(MSC)寻找此模型的最大似然估计参数,对异常情况具有很强的鲁棒性。在Yale及ORL人脸数据库的实验结果表明了该方法对于人脸模糊、光照及表情变化等的有效性及鲁棒性。

Sparse coding （SRC） is an effective method for face recognition. The detected image is represented as a sparse linear combination of a set of training samples, the accuracy represented by L2 or L1 norm residue to measure. This model assumes that the encoding residual items Ganssian or Laplace distribution. In fact it can not be very accurate description of coding error rate. In this paper, a new sparse coding method is proposed to establish a model of constrained regression problems. SRC for finding the maximum likelihood estimation parameters of this model has a strong robustness to abnormal situations, namely MSC. The experimental results on Yale and ORL database show the effectiveness and robustness of the method for the human face blurred, illumination and expression changes.