基于低秩表示的鲁棒特征抽取和分类方法研究

项目名称： 基于低秩表示的鲁棒特征抽取和分类方法研究

项目编号： No.61503188

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 自动化技术、计算机技术

项目作者： 陈燚

作者单位： 南京师范大学

项目金额： 21万元

中文摘要： 在非受控环境下，如何有效消除环境噪声影响是生物特征识别技术面临的难点。近年的低秩分解理论揭示，许多实际观测量都可以归结为一个低秩分量和稀疏噪声分量相加的模式。借助矩阵的低秩分解理论，可以从噪声或污染数据中恢复原始数据信息。低秩分解理论通常假设噪声数据服从拉普拉斯分布或者高斯分布，在复杂的现实场景中往往是不成立的。本项目提出基于低秩分解理论的噪声自适应的鲁棒特征提取方法和鲁棒分类方法。提出的方法能根据不同的场景自适应地过滤掉观测数据中的噪声污染，具有较强的鲁棒性，有效地解决了生物特征识别在非受控环境下的噪声污染问题，具有较高的理论价值和应用价值。本项目研究将人类感知图像的稀疏性机制、低秩表示理论与流形学习的研究结合起来，丰富和发展了模式识别的理论体系；在技术上设计出更具鲁棒性和鉴别能力的特征抽取和分类算法，为非受控环境下的鲁棒生物特征识别提供更好的技术支撑。

中文关键词： 特征提取；模式分类；低秩表示；生物特征识别；稀疏表示

英文摘要： How to alleviate the impacts of noises in an uncontrolled environment is a key problem of biometrics. Recent studies on low rank decomposition theories (LRDT) reveal that many observations can be intrinsically expressed as a sum of a low-rank component and a sparse component. The clean data can be recovered from observations corrupted by gross noises or outliers through the LRDT. According to the assumption of LRDT, the noises of observations follow the Laplacian distribution or Gaussian distribution, which does not hold in many real applications. Based on LRDT, we developed noise adaptive robust feature extraction and robust classifications methods. The proposed methods can automatically estimate the distribution of the environment noises and filter the noises adaptively. Our methods address the noise issue in an uncontrolled environment efficiently due to their robustness to the noises. The research will produce significant values both in theoretic and application areas. The projection combines sparse mechanisms of human perception image, LRDT and manifold learning together, enriching the theoretical systems of pattern recognition and feature extraction techniques. Designing the feature extraction and classification algorithms with more robustness and discrimination ability is to provide better technical support for biometrics in an uncontrolled environment.

英文关键词： feature extraction;pattern classification;low rank representation;biometrics;sparse representation