结构化判别字典学习方法及其应用研究

Discriminative Dictionary Learning (DDL) is one of the frontier topics in the fields of machine learning and pattern recognition. Up to now, researcheres have developed a few DDL approaches under the sparse representation framework. However, most of them fail to generate high accurate classification in many realworld situations. Meanwhile, training their learning models may cost a large amount of computation time. To address these issues, this project targets at developing a new class of DDL models in a distinguishing way of structural modeling. In this process, theoretical analyses behind the modeling will be conducted to construct and optimize the DDL models. Specifically, the research work will mainly focus on developing the following learning models, inclduing structured sparse DDL model, self-taught DDL model, and multi-view DDL model. Studying with these models, a new class of DDL algorithms will be finally proposed. To this end, we will study how to derive and formulate structured constraints from supervised information and clusters, so as to improve the discriminative ability of the DDL models. In addition, to embed an effective and efficient classifier into the DDL models to be constructed, we will study how to develop a discriminative regression model with compact form under the large margin framework for multi-class classification. It will be proved that, this discriminative regression model is equivalent to the soft large margin support vector machine, and thus shares the equal power for data classification. Meanwhile, due to its compact form, the model has a low complexity and thus will facilitate its training process. Finally, we will construct a unified function expression model for multiple kinds of norms of vector and matrix. This will eventually provide a theoretical tool for efficiently solving the constructed DDL models.. As for the application issues, the proposed algorithms will be applied to two special tasks of image segmentation. One is to segment object-of-interest with exemplars, and the other is to segment image into super-pixels. To this end, an experimental platform will be constructed to evaluate the proposed DDL algorithms and the image segmentation algorithms. Upon this platform, the related algorithms will be analyzed comprehensively to improve the segmentation quality and reduce the computation time. As a result, the combination between DDL and image segmentation here could offer a new vision for the practice of image segmentation and other tasks of image understanding.

判别字典学习是机器学习和模式识别中的一个前沿性课题。现有算法在分类性能、模型训练等方面还不能满足现实的应用需求。本项目拟研究结构化判别字典学习建模方法和算法理论，形成包含结构化稀疏判别字典学习、自导判别字典学习和多视角判别字典学习等在内的字典学习算法。在学习模型构建中，基于监督信息和聚类信息，项目将引入结构化约束来提高字典学习模型的分类判别能力；同时，研究形式紧凑的最大间隔多类分类判别回归模型的构建方法，并将其作为分类器嵌入到判别字典学习模型之中，保证分类性能的同时降低模型复杂度。此外，项目还将构建能够统一多种范数的函数表示模型，以此为基础实现判别字典学习模型的高效求解。.　　另外，对本项目提出的算法，拟应用于基于样例的图像感兴趣目标分割和图像超像素分割。为此，项目拟开发一个图像分割实验平台，在此平台上验证和分析项目提出的学习算法，提高图像分割的精度和速度，进而促进学习算法的应用研究。

本项目研究结构化判别字典构造方法以及相关基础性机器学习问题。研究内容主要包含判别回归模型、稀疏学习模型、结构化稀疏判别字典学习方法、自导字典学习方法和多视图字典学习方法。同时，将相关研究方法或思想应用于图像分割。. 自项目开展以来，项目组按照研究目标、研究内容、研究计划部署和预期研究成果要求，对拟解决的关键科学问题进行了深入分析，并提出了相应的解决方法。在项目执行过程中，对相关研究内容，不仅给出了相应的理论分析，也给出了完整的算法流程，并对算法的性能进行了实验验证。项目执行情况良好，并因此获得了较好的研究成果。. 取得的重要成果如下：（1）提出基于最大间隔准则的判别最小二乘回归模型和目标重定位最小二乘回归模型，其模型形式紧凑，利用嵌入和扩展；（2）提出一个核稀疏学习框架，可普适于多种不同的核构造；（3）提出一种树形判别字典构造方法，可适用于多模式分布的变化数据；（4）提出一种结构化稀疏非负矩阵分解方法，通过图拉普拉斯光滑和丰度系数稀疏表示提高了字典（端元）的表示能力；（5）提出一种基于样本自表示的快速子集选择方法，可选择出信息量丰富、价值大的样本；（6）提出一种多视图耦合字典学习方法，实现了不同视图间的信息共享；（7）提出一个端对端跨模态深度双向表示模型，在度量学习目标的驱动下，所学各卷积层的滤波器（字典）更具判别能力；（8）提出基于邻域传播的精细结构目标分割方法，提高了目标分割的精度。总之，针对传统算法中的不足，项目构建了一些新模型和新方法，具有重要的理论意义和实际应用价值。. 本项目共发表25篇学术论文。其中，部分研究成果已发表于IJCV (1篇)、IEEE T-NNLS (2篇)、IEEE T-MM (2篇)、JPRS(1篇)、Pattern Recognition (1篇)等本领域国际顶级杂志或AAAI、CIKM等本领域国际知名会议上。