复杂场景下非合作目标鲁棒识别方法研究

项目名称： 复杂场景下非合作目标鲁棒识别方法研究

项目编号： No.61472379

项目类型： 面上项目

立项/批准年度： 2015

项目学科： 计算机科学学科

项目作者： 庄连生

作者单位： 中国科学技术大学

项目金额： 83万元

中文摘要： 以压缩感知理论为基础、以人脸为研究对象，本课题主要针对复杂场景下非合作目标鲁棒识别的理论和方法开展研究，解决图像序列的多任务学习识别、保持图像线性子空间结构的光照归一化、姿态和表情联合建模这三个关键科学问题，建立基于GPU架构的鲁棒物体识别演示验证系统，旨在把稀疏表示鲁棒物体识别技术从可控或半可控场景拓展到非可控场景，期望建立较完善的以压缩感知理论为基础、面向非合作目标、适合复杂场景的鲁棒物体识别相关理论和方法，在物体识别领域产生较大的国际影响。 本课题研究的主要贡献包括:(1)提出基于多任务学习的结构稀疏表示联合识别框架;(2)提出基于字典共享的稀疏光照迁移方法，以及高效的稀疏光照字典学习算法；（3）提出基于贝叶斯网络的人脸分块表示模型和分块图像对齐算法；（4）建立基于GPU架构的鲁棒物体识别验证平台，为新理论和新方法的性能验证提供平台。

中文关键词： 鲁棒物体识别；稀疏表示；字典学习；多任务学习

英文摘要： Based on compressive sensing theory, this project aims at figuring out new theories and methods on robust non-cooperative object recognition in complex scene, by taking face as a research example. Three key issues will be addressed: multi-task learning based object recognition in image series, linear-structure preserving illumination normalization, and jointly modeling pose and expression. A GPU-based prototype system on robust object recognition will be built to verify the above theories and methods. Our goal is to extend the study of sparse representation based robust object recognition from a controlled or partly controlled scene to a uncontrolled scene. We hope to establish a relatively complete theory and method for robust object recognition based on compressive sensing theory, which is also suitable to non-cooperative object recognition in complex scene. This research project will have an important impact on object recognition over the world. The main contributions of our research include: (1) Propose a multi-task learning based object recognition framework via structured sparse coding. (2) Propose a dictionary sharing sparse illumination transfering method, and an efficient dictionary learning method for sparse illumination transfering. (3) Propose a part based face model by Bayesian network, and a new method for part based face alignment. (4) Build a GPU-based system for robust object recogniton to provide a platform for validating the performance of new theories and methods.

英文关键词： Robust Object Recognition;Sparse Representation;Dictionary Learning;Multi-Task Learning