项目名称: 基于流形理论和稀疏表示的低质量图像人脸识别算法研究
项目编号: No.61203241
项目类型: 青年科学基金项目
立项/批准年度: 2013
项目学科: 自动化学科
项目作者: 樊明宇
作者单位: 温州大学
项目金额: 25万元
中文摘要: 随着公共安全方面需要的增长,更多的人脸识别应用需要在图像质量不可控的情况下使用。然而低质量图像给人脸识别带来了非常大的挑战,现有的识别系统在这样的情况下性能下降非常快,不能达到实用水平。本项目拟针对低质量图像的人脸识别问题,从机器学习的角度开展创新性的研究工作,力求实现在低质量图像人脸识别问题中若干关键问题的突破性进展。项目的难点体现在:1)如何在单个像素信息不可靠的情况下提取低分辨率高噪声人脸图像整体的本质特征,2)如何实现不减弱判别性能的人脸图像分割采样方法并将其用于解决人脸被遮挡情况下的识别问题,和3)如何利用稀疏编码中的判别信息实现快速、对噪声和遮挡鲁棒的分类算法。根据本项目的难点,我们的具体研究子目标包括1)新型的流形学习理论用于低质量人脸图像的本质特征提取,2)一种增量、可在线更新的子空间优化模型,和3)基于人脸局部特征和全局特征有效结合的集成分类算法用于遮挡情况下人脸识别。
中文关键词: 模式识别;人脸识别;流形学习;图像分类;
英文摘要: More and more face recognition systems are applied under unconstrained conditions for public security. This is a great challenge to classical face recognition algorithms. The experiments indicate that the performances of existing face recognition systems decline rapidly under the unconstrained conditions. In this project, we plan to resolve the unconstrained face recognition problem from the prospective of machine learning. The key problems in this project are 1) how to extract the most essential features from the low-quality images when the information of a single pixel is no longer reliable, 2) how to segment the face images without tempering the discriminative power and apply the segment method to images with occluded faces, and 3) how to apply the sparse representation of data to construct a fast, robust classifier. The concrete research of this project include three folds: 1) a new and effective manifold learning theory for the feature extraction of low-quality images, 2) a fast and incremental subspace optimization model for our manifold learning algorithm and 3) a ensemble classifier based on the global and local features for occluded face recognition.
英文关键词: Pattern recognition;Face recognition;Manifold learning;Image classification;