Face Recognition Via Sparse Representation
Face Recognition Via Sparse Representation
复制标题
DOI：
10.1002/047134608x.w8276
复制
发表时间：
2015-09
期刊：
影响因子：
--
通讯作者：
Meng Yang
中科院分区：
文献类型：
--
作者：
Meng Yang
近年来，稀疏表示在包括人脸识别在内的各种分类任务中得到了快速发展。基于稀疏表示的分类（SRC）在FR上的成功极大地推动了基于稀疏表示的分类技术的研究。然而，仍有许多问题有待进一步解决。例如，l0或l1范数稀疏性在其中的作用是什么？如何设计一个鲁棒的表示保真度项来处理各种离群值？如何提取有效的特征以提高SRC的准确性和效率？在这篇文章中，我们旨在通过三个方面来回答这些问题：稀疏系数的作用：一些作品质疑稀疏系数的作用。特别地，已经针对FR提出了基于协作表示的分类，其使用l2范数来正则化编码系数，具有与SRC相似的精度，但速度快得多。然而，也有一些工作，捍卫稀疏的编码系数时，使用适当的字典。表示残差的正则化：已经开发了具有有效表示项的鲁棒模型来处理人脸图像中的各种离群值。例如，假设编码残差和编码系数独立同分布，提出了基于最大后验估计的正则化鲁棒编码。此外，还提出了利用图像结构信息的结构化表示模型。扩展稀疏表示模型：FR的扩展稀疏模型包括未对准鲁棒模型、欠采样模型和多特征模型。这些通过稀疏表示的FR方法已经取得了非常有希望的结果。关键词：人脸识别;稀疏表示
Recent years have witnessed rapid developments of sparse representation in various classification tasks, including face recognition (FR). The success of sparse representation-based classification (SRC) on FR greatly boosts the research of sparsity-based classification techniques. However, many problems are still pending to be further addressed. For example, what is the role of l0 or l1 norm sparsity in it? How to design a robust representation fidelity term to handle various outliers? How to extract effective features to improve the accuracy and efficiency of SRC? In this article, we aim to answer these questions via three aspects: Role of sparse coefficient: Several works have questioned the role of sparse coefficients. Especially, a collaborative representation-based classification has been proposed for FR, which use l2-norm to regularize the coding coefficient, with similar accuracy to SRC but much faster speed. However, there is also some work that defends the sparsity of coding coefficients when an appropriate dictionary is used. Regularization on the representation residual: Robust models with effective representation terms have been developed to handle various outliers in face images. For instance, by assuming the coding residual and the coding coefficient are respectively independent and identically distributed, regularized robust coding was proposed based on maximum a posterior estimation. Moreover, structured representation models were also proposed to make use of the image structure information. Extended sparse representation model: Extended sparse models for FR include misalignment robust models, undersampled models, and multifeature models. These FR approaches via sparse representation have achieved very promising results. Keywords: face recognition; sparse representation