基于图理论的半监督图像分割与应用研究

As one of the most important and typical problems in image processing and computer vision fields, image segmentation by which the pixel is divided into visually meaningful regions is the basic premise in image vision analysis and pattern recognition. In recent years, as a branch of the semi-supervised learning, graph-based semi-supervised learning has drawn more and more attention since graph-based semi-supervised learning is very important to improve the performance of machine learning system. In this project, image segmentation algorithm of graph theory based semi-supervised learning has been explored by using graph theory, Gabor transform, spectral structure tensor clustering combined with multi-scale analysis as a tool. Based on a comprehensive data analysis, we shall resort to the idea such as manifold learning and semi-supervised learning, combine local and global topological structure, consider the image data set, construct the adjacency matrix structure sample, investigate the effect of redundant features and noise as well as the dimension of the distance measure on graphs, and random subspace method of random sampling for the feature subset of image, In addition, we shall establish a local adaptive graph model which is suitable for color image segmentation. Further, a semi-supervised dimensionality reduction method of integration domain transform global and local topology will be proposed. In order to improve the computational efficiency of graph, we shall study image segmentation method for semi-supervised learning based on multi-scale analysis. At the same time，some researched will be carried out for image segmentation，such as sparse similarity matrix calculation, graph based structure and constraints, the semi-supervised feature selection method, semi-supervised manifold learning algorithm, constrained energy function, spectral clustering method. By means of the exhaustive research for this project, the discipline such as image processing, computer vision, etc will be further promoted and developed.

图像分割是图像处理中重要的基础环节，它将像素分隔成有意义的区域。基于图的半监督学习是半监督学习领域的主流方法之一。本项目以图理论、域变换、谱聚类以及多尺度的结构张量分析为工具，探讨基于图理论的半监督学习的图像分割算法。从数据分析的角度，充分将流形学习和半监督学习的思想相结合，综合考虑图像数据的局部和全局拓扑结构，构造样本的邻接矩阵，探讨冗余特征以及维数对图上距离测度的影响，结合特征子集进行随机采样的随机子空间方法，建立一种适用于图像分割的局部自适应图模型；研究一种基于域变换的整合全局和局部拓扑结构的半监督维数约减方法。以提高计算效率为目的，研究多尺度图理论的半监督学习的图像分割算法。同时对稀疏相似矩阵计算、结构和约束保持的半监督图像特征选择、约束能量函数的半监督流形学习算法以及谱聚类的图像分割算法等方面展开研究。本项目的深入研究，将对图像处理与计算机视觉学科等的发展起到一定的推动作用。

图像分割是数字图像分析中的最为困难和也是最重要步骤，它既是对所有图像预处理效果的一个检验，也是后续进行图像分析与解译的需要。分割是将图像分成各具特性的区域并提取出感兴趣目标的过程，该技术广泛应用于医学、军事、体育、遥感、智能交通、产品检测、网络以及计算机视觉等领域。本课题以金石拓片图像、植物叶片图像、交通信号图像分割为应用背景，针对碑帖拓片图像的低对比度特征，我们探讨了五种不同的前处理方法对碑帖图像的去噪处理研究，分析了全局和局部阈值化处理方法在数字碑帖图像文字分割中的应用；采用深度学习特征提取等方法对叶片图像特征进行提取, 解决多种植物叶片图像特征的整合问题,以获得高质量的植物叶片图像综合特征，开发了交通信号图像的分割提取与分类研究，为进一步探索有效的数字图像分割、字符识别、植物分类、交通图像自动识别提供进一步的理论依据。对石刻碑帖的数字化拓片的研究，对继承和发扬传统书法文化以及研究中国古代史具有十分重要的意义。同时在植物叶片分类、交通信号处理等的研究上，将对我国正大力倡导的数字化园林植物研究和数字农业、智能交通等起到一定的推动作用。