视觉识别中的实用鲁棒回归技术研究

项目名称： 视觉识别中的实用鲁棒回归技术研究

项目编号： No.61502081

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 其他

项目作者： 沈复民

作者单位： 电子科技大学

项目金额： 20万元

中文摘要： 利用线性回归产生的回归残差作为判别依据的识别技术是当前模式识别特别是视觉识别领域的一个热门研究课题。这类技术假设测试样本存在于同类的样本展开的线性子空间内，即测试样本可由同类样本来线性表示。这类方法虽然简单，但在很多情况下仅利用原始特征即可取得很好的识别效果。然而，样本（如图像）的特征中往往存在大量的噪声，这使得线性子空间的假设失效，从而严重影响表示系数的回归估计，使得识别率下降。本项目以人脸识别为例，拟重点研究实用的鲁棒回归算法，以期可以准确的检测出图像中的噪声孤立点，从而做出准确的识别。本项研究的关键在于实用鲁棒回归算法的高准确性以及高计算效率。本研究的主要创新之处在于：1）引入最小截平方和回归框架，并研究快速的近似算法；2）提出利用最小化无穷范数来快速检测回归中的孤立点；3）提出一个基于局部性限制的鲁棒回归技术。

中文关键词： 视觉识别；鲁棒回归；线性表示；人脸识别

英文摘要： Recognition based on the residuals of linear regression model has been a hot research topic in recent visual recognition community. These methods are based on the assumption that a high-dimensional probe sample lies on a low-dimensional subspace spanned by the training samples of the same subject. Despite their simplicities, these approaches can achieve excellent results with features of raw pixels in some situations. In practice, however, various potential noises of the features make the linear subspace assumption not hold. The estimation of the coefficients are largely biased by the noisy points (outliers), which deteriorates the recognition accuracies. Take face recognition as an example, this research project will focus on robust regression, which can effectively detect the noisy pixels in the images. Based on that, more accurate coefficients are estimated, which helps to yield better recognition. The key points of this research are the high accuracy and high computational efficiency. The novelty of this project includes: 1) the least trimmed sum of squares (LTS) regression is introduced to face recognition; 2) the L∞-Minimization is introduced to detect outliers; 3) a locality constrained representation and classification method is proposed.

英文关键词： visual recognition;robust regression;linear representation;face recognition