基于张量流形学习的图像特征提取方法研究

Image feature extraction, as an important research topic in pattern recognition and computer vision, has been paid considerable attention in both academic and industrial domains. A critical problem in this research topic is to design a type of image representation which is both expressive and robust to appearance variantions. Oriented to the two important issues of image model and the underlying structure of image, this project would investigate the nonlinear feature extracted from image by integrating tensor model and manifold learning. Particularly, the following issues would be addressed: (1)Designing manifold learning methods in tensor space for image dimensionality reduction;(2)For the facility of feature extraction, exploring feature space with multiple linear projection to extimate the underlying mapping of manifold embedding;(3)Extending tensor manifold learning to the case of mutiple-manifold to further strengthen the ability of feature. This project will perform image feature extraction with a new pesective of tensor manifold, and provides new thougths for building practical image feature extraction systems. The research can produce inchievement as the theoretical and technological support for the applications of image understanding, biometric recognition, intelligent surveilance and multimedia retrieval etc. These achievements can also promote the development and application of intelligent image information systems in both manufacturing industry and daily life.

图像特征提取一直是模式识别和计算机视觉领域具有重要理论价值和广阔应用前景的研究课题，研究该课题的关键任务之一是设计表达能力强、对表观变化鲁棒的图像表示方法。针对这一任务，本项目从图像的数据模型和潜在模式结构这两个关键因素出发，融合张量模型和流形学习理论，研究面向图像的非线性特征提取方法。具体研究内容包括：（1）研究张量空间的流形嵌入方法对图像进行降维；（2）设计多重线性投影对流形嵌入映射进行近似来得到图像特征空间；（3）探索图像张量的多流形学习方法，增强对复杂结构图像的特征提取能力。本研究将建立张量模型和流形学习之间的桥梁，在图像特征提取中融入张量流形这一新的观点，为建立实用的图像特征提取系统和方法提供新的思路。本项目的研究成果可为图像理解、生物特征识别、智能监控和多媒体数据检索等应用领域提供理论基础和技术支撑，推动智能化图像信息处理技术在生产生活中的应用和发展。

该项目研究了面向复杂流形结构的图像特征提取。研究内容包括：1. 基于张量流形嵌入与多重线性投影的图像降维与特征提取方法；2. 领域自适应情形下的非线性图像特征提取方法；3. 零样本图像识别中的图像潜在结构挖掘与特征提取方法。取得的研究成果包括：1. 利用图像张量流形建模提取图像识别中的领域不变特征；2.图像张量模型下基于矩阵回归的图像特征刻画与图像超分辨率；3. 基于图像张量流形模型和自编解码的零训练样本图像特征提取新方法。该项目探索了当前图像识别领域前沿应用中的图像特征提取问题，其研究成果对模式识别领域某些问题的解决具有促进作用。