基于GIST特征和流形学习的特征抽取方法的研究

发布时间：2018-04-25 11:30

本文选题：Gist特征 + 显著性计算　；参考：《吉林大学》2015年博士论文

【摘要】：本文对基于邻接图的流形学习降维算法，基于梯度方向直方图的人脸特征提取算法和gist特征抽取算法进行了深入研究，，分别对基于邻接图的流形学习降维算法，基于梯度方向直方图的人脸特征提取算法和gist特征抽取算法提出了改进算法，并基于这些改进算法提出了新的人脸识别方法和建筑物识别方法。经典的k近邻(k-nearest neighbors)邻接图构造方法，通过计算样本向量的k个近邻来构造流形学习算法的邻接图。这种方法没有在构建邻接图的时候考虑样本的原始结构信息，针对这一缺陷，笔者提出了基于样本对应列的邻接图(CorrespondingColumnsBasedGraph, CCG)构造方法。笔者在基于样本对应列的邻接图构造方法的基础上提出了基于样本对应块的邻接图(CorrespondingblockBasedGraph,CBG)构造方法。实验证明，基于样本对应块的邻接图对人脸识别中遇到的非均匀光照具有一定的鲁棒性。经典的k近邻邻接图构造方法在构建邻接图时，需要凭借经验指定近邻参数k。针对这一问题，笔者在CCG和CBG的基础上提出了基于样本内部结构的邻接图构造(Samples’Inner Structure Based Graph, SISG)方法。笔者提出了用于人脸识别的局部敏感梯度方向直方图(Locality Sensitive Histogramsof Oriented Gradients, LSHOG)。该梯度直方图在提取人脸特征向量的时候体现了人脸图片的二维结构信息和全局信息，因此局部敏感梯度方向直方图对于人脸上的遮挡和非均匀光照有一定的鲁棒性。本文在传统的gist特征的抽取方法的基础上提出了子区域多尺度gist特征提取方法。相比于原有的gist特征提取方法，本文提出的子区域多尺度gist特征提取方法对于建筑物图片中的非均匀光照具有较高的鲁棒性。笔者将子区域多尺度gist特征提取方法和基于样本内部结构邻接图的流形学习算法相结合提出了一种新的建筑物识别方法：基于子区域多尺度gist特征和样本内部结构邻接图降维算法的建筑物识别方法。
[Abstract]:In this paper, manifold learning dimensionality reduction algorithm based on adjacent graph, face feature extraction algorithm based on gradient direction histogram and gist feature extraction algorithm are studied in detail. Based on gradient direction histogram and gist feature extraction algorithm, an improved face recognition method and a new building recognition method are proposed. The classical k-nearest neighbor adjacent graph construction method is used to construct the adjacent graph of the manifold learning algorithm by computing k nearest neighbors of the sample vector. This method does not take the original structure information of samples into account when constructing adjacent graphs. In view of this defect, a method of constructing adjacent graphs based on Corresponding columns (CCGs) based on sample corresponding columns is proposed in this paper. Based on the method of constructing adjacent graph based on sample correspondence column, a method of constructing adjacent graph based on Correspondingblock based CBG (CBG) based on sample corresponding block is proposed in this paper. The experimental results show that the adjacent graph based on the corresponding block of samples is robust to the non-uniform illumination in face recognition. The classical construction method of k-nearest neighbor graph needs to specify the nearest neighbor parameter k by experience when constructing the adjacent graph. In order to solve this problem, based on CCG and CBG, the method of constructing adjacent graphs based on internal structure of samples is proposed, which is called Inner Structure Based Graph (SISG). The local Sensitive Histogramsof Oriented gradient histogram for face recognition is presented in this paper. The gradient histogram presents the two-dimensional structure and global information of the face image when extracting the face feature vector, so the local sensitive gradient direction histogram is robust to the occlusion and non-uniform illumination of the human face. Based on the traditional gist feature extraction method, a multi-scale subregion gist feature extraction method is proposed in this paper. Compared with the original gist feature extraction method, the multi-scale gist feature extraction method proposed in this paper is robust to non-uniform illumination in building images. In this paper, a new building recognition method is proposed by combining subregion multi-scale gist feature extraction method and manifold learning algorithm based on adjacent graph of sample interior structure: based on sub-region multi-scale gist feature and sample interior knot. Building recognition method based on dimension reduction algorithm of adjacent graph.
【学位授予单位】：吉林大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【参考文献】