SGER: Explorations of Robust Image Classification

SGER：鲁棒图像分类的探索

• 批准号: 0849292
• 负责人: Yi Ma
• 金额: $ 10万
• 依托单位: University of Illinois at Urbana-Champaign
• 依托单位国家: 美国
• 项目类别: Standard Grant
• 财政年份: 2008
• 资助国家: 美国
• 起止时间: 2008-09-15 至 2009-08-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=0849292&HistoricalAwards=false
• 关键词: SGER; Explorations; Robust; Image; Classification

One of the fundamental problems in computer vision is to make a machine automatically recognize an object in a query image based on previously seen examples. This problem becomes particularly difficult when the machine is trying to recognize an object among several with rather similar appearances, and the object is partially occluded or disguised. This is often the case with human face recognition. This project aims to explore new mathematical tools from sparse representation in signal processing, by casting robust face recognition as a (sparse) error correction problem. Recently, sparse error correction based on minimizing the 1-norm has seen great success in signal processing. This project will investigate its potential in image-based object recognition despite occlusion or corruption, especially for human faces. Preliminary experimental results have shown good promise of this new approach. In its one-year span, this project aims to study the special geometric and statistical models and problems associated with human face recognition and hopes to develop even more robust and scalable face recognition algorithms. To verify the results, a prototype face recognition system will be developed, with an emphasis on theoretical and algorithmic progress. All the results will be available at a public website: http://perception.csl.uiuc.edu/recognition/Home.html

计算机视觉的基本问题之一是使机器根据先前看到的示例自动识别查询图像中的对象。当机器试图在几个外观相当相似的物体中识别一个物体，并且该物体部分被遮挡或伪装时，这个问题变得特别困难。这通常是人脸识别的情况。该项目旨在通过将鲁棒人脸识别作为（稀疏）纠错问题，探索信号处理中稀疏表示的新数学工具。近年来，基于最小化1范数的稀疏纠错在信号处理中取得了很大的成功。该项目将研究其在基于图像的物体识别方面的潜力，尽管有遮挡或损坏，特别是对人脸。初步实验结果表明，该方法具有良好的应用前景。在为期一年的时间里，该项目旨在研究与人脸识别相关的特殊几何和统计模型和问题，并希望开发出更强大和可扩展的人脸识别算法。为了验证结果，将开发一个原型人脸识别系统，重点是理论和算法的进步。所有结果将在一个公共网站上公布：http://perception.csl.uiuc.edu/recognition/Home.html