基于噪声建模的统一特征提取模型、理论、方法与应用

The main aim of this project is to construct a unified framework for feature extraction through modeling the intrinsic noises mixed in data. The new frame work mainly aims to solve the feature extraction problem on data with complicated noises and missing components. The existing methods for feature extraction model the problem by virtue of L2- or L1-norm errors, which are intrinsically effective only on data embedded with Gaussian or Laplacian noises. Different from the previous investigations, we will model the noises mixed in data based on their intrinsic structures, and then utilize the maximization likelihood estimation method to solve the proposed model. The propose method is hopeful to achieve both the goals of feature extraction and noise learning, simultaneously. The proposed models and methods incline to significantly extend the available range of feature extraction, and tend to achieve the goals of both adaptive noise learning and robust feature extraction on data with complicated noises and missing components. All of the above investigations are not only theoretically meaningful for clarifying the intrinsic principle underneath the feature extraction problem and revealing the essential effective mechanism underlying feature extraction methods, but also potentially useful on practical applications such as face recognition and small target recognition in a single infrared image.

本项目将基于对复杂噪声进行建模的手段建立统一的特征提取研究框架，主要针对具有复杂噪声结构与缺失信息数据的特征提取问题展开研究。传统特征提取方法主要基于L2与L1误差范数对问题进行建模，其有效性本质限定于混有高斯与拉普拉斯分布噪声的数据。本项目将通过对噪声结构分布进行建模，并进而利用最大似然方法对模型进行求解，从而同时完成数据本质特征提取与复杂噪声结构学习两个目标。所提思想将本质扩展特征提取方法的有效应用范围，进而对蕴含有复杂噪声结构与缺失信息的数据进行有效的自适应性噪声学习与稳健性特征提取。所获结果不仅对于阐明特征提取问题的实质规律，揭示特征提取方法的内在有效性机制方面具有重要理论意义，也在人脸识别与红外图像小目标识别等实际问题中拥有广阔的应用前景。

项目原计划目标为从模型、理论、方法与应用层面研究具有复杂噪声结构数据的鲁棒性机器学习问题。经过四年的执行，项目组圆满完成原项目计划，并对原项目任务进行了显著性扩展。项目成果共包括59篇论文，其中SCI文章35篇，IEEE Trans论文14篇，期刊影响因子大于4的文章17篇，影响因子大于2的论文29篇；CCF A类会议18篇，CCF B类会议5篇；在人工智能学会通讯撰写科普报告一篇；申请国家发明专利4项（待授权）。项目主要成果介绍如下：针对复杂噪声结构与缺失信息数据问题原创性地提出“基于数据驱动的误差建模”与“基于认知机理的自步学习”两类机器学习建模新型机理，本质性扩展了原定项目方案，形成了针对蕴含复杂噪声信号重建问题与蕴含干扰标记/无标记/弱标记目标识别问题的新型处理策略，所提鲁棒学习方法论不仅在原项目计划的人脸识别与红外小目标检测问题上获得成功应用，且将其有效应用于高光谱去噪，稳健性视频压缩，生物化学结构检测，CT图像增强，鲁棒多曝光率图像融合，多媒体事件检测，弱监督视频概念检测，视频语义搜索，半监督示例标定等多项实际问题中。项目成果支持CMU信息媒体实验室与香港中文大学在TRECVID与IEEE GRSS竞赛中取得优异成绩。