【作者】 刘悦；

【导师】 谢红；

【作者基本信息】 哈尔滨工程大学 ， 通信与信息系统， 2015， 博士

【摘要】 在人类认知和理解周围环境的过程中,图像成为视觉上一种最直观、最容易的获取方式,如何从图像中获得重要信息成为对图像分析和理解的关键。图像特征提取方法的出现,有效的解决了这一关键性问题。对图像信息的描述常借助于一些称为目标特征的描述符来进行,目标特征代表了目标区域的特性。不同的图像类型具有不同的目标特征,因此具有针对性的对不同的图像类型研究多种与之相适应的特征提取算法对于有效进行图像识别与分类具有重要意义。本文的重点是对图像特征提取算法的研究及应用,针对图像中待识别与检测特征的不同,围绕基于灰度统计特性的特征提取算法、基于形态成分分析的特征提取算法、基于核学习的特征提取算法以及它们在相关图像中的典型应用展开研究。论文的主要工作及创新性表现包含以下几个方面:特征提取算法的性能优劣与原始图像的质量有着密切的关系,噪声是影响图像质量的主要因素,在某些不理想的情况下有必要在分析图像之前进行预处理从而消除噪声。针对这一问题,论文在介绍基于灰度统计特性的特征提取方法之前,先对频域滤波方法和空域滤波方法进行了介绍,阐述了小波(Wavelet)、轮廓波(Contourlet)、非下采样轮廓波(Non-subsampled Contourlet)、非局部均值(Non-local means)方法以及中值滤波算法的原理,探究它们的数学模型。在此基础上,联合空域及频域滤波方法提出基于变换域的双重滤波去噪算法,通过对不同的噪声特点采用适合的处理方法,较好的去除了图像中的高斯噪声和脉冲噪声,突出图像中的目标特征使后续边缘检测更加准确。其次,针对图像处理中特征提取这一模块提出了基于灰度统计特性的特征提取算法。该算法针对直方图具有单峰形态的图像进行研究,将Rosin方法作为Canny中的阈值设定方法来实现边缘检测。考虑到一维Rosin算法在处理此类图像时不能够充分利用图像局部结构信息这一缺陷,将传统Rosin方法推广至二维,利用像素点本身和其邻域的局部结构信息,建立图像灰度的二维直方图,利用二维直方图找到适合提取图像特征的阈值,从而为Canny算法中的阈值选取获得一种自适应的方法。在仿真实验中,采用具有单峰形态直方图特性的图像验证了算法的有效性,同时将其应用于道路裂纹图像的边缘检测并获得理想的效果。再次,提出了基于形态成分分析的特征提取算法,针对具有复杂背景的图像难以获取主要特征的问题,构造基于形态成分分析的活动轮廓模型。首先采用不同的字典对图像的不同成分进行稀疏表示,从而分离图像不同成分的特征,再利用活动轮廓模型寻找到图像中所需内容主要轮廓特征,仿真实验证明提出的算法实现了对图像中不同纹理特征的分离,达到了对特征识别及检测的目的。最后,提出了基于多数据相关核Fisher判别分析的图像特征提取算法。针对传统核Fisher判别分析方法对于包含多源、异构数据的图像识别适应性较差的问题,将核优化思想应用于核Fisher判别分析中,介绍了数据相关核及多核方法,并联合这两种方法对传统单核函数进行改进以及参数优化,使得传统核Fisher判别分析方法对于包含多源、异构数据的图像具有更好的特征提取效果。通过人脸图像识别的应用验证了提出方法的有效性。更多还原

【Abstract】 In the process of human cognition and understanding of the surrounding environment,image becomes most intuitive and convenient means for human visual.So,the key of image analysis and understanding is how to get important information from the image.The emergence of image feature extraction solved the critical question effectively.The image information can be described usually by means of a descriptor for target characteristics,in which target feature represent the characteristics of the target area.Different types of image have the different target characteristics,so develop effective feature extraction algorithm which aim at different types of image is very important.Aimed at the feature extraction research and application algorithm of image,several new feature extraction algorithms were proposed,including based on gray statistical properties,signal processing and learning algorithm.The specific contents are as follows:Since the performance of edge detection algorithm is closely related with the quality of the original image,noise is the major interference factor,which means it needs preprocessing before image analysis,this paper introduced filtering method first,in order to get better effect in the chapter three.The second chapter first discusses some basis methods about denosing theory,which mainly includes Wavelet,Contourlet,Non-subsample Contourlet,Non-local means and median filter.And on this basis we combine the spatial domain and frequency domain to propose a NSCT image de-noising method with double filtering,which adopt different method to remove different noise.The simulation results show that,it can highlight the characteristics of image so that getting better results of edge detection.Secondly,for the feature extraction module of image processing,the feature extraction methods based on grey statistical property were proposed.This algorithm is aimed specially at unimodal pattern histogram of image.We combined the Canny edge detection algorithm and Rosin algorithm,then extended the traditional Rosin method to 2-D based on the defect that the traditional one dimensional Rosin algorithm cannot make full use of local structure information.The pixels itself and its neighborhood of local structure information are used to establish the 2-D histogram of image.Using the peak and the bottom point of histogram tostructure the whole angle rotation threshold selection plane and search the maximum value of space distance between histogram and every threshold plane,then got an adaptive threshold selection method.The test result show that the effectiveness of algorithm and expand the application scope of the canny algorithm.Thirdly,the feature extraction algorithm based on morphological component analysis(MCA)was proposed,by the implementation of MCA,the generated two image series of smooth piecewise edges and textures components can be described by suitable dictionaries,and the combination of smooth piecewise part and active contour model is achieved.Experimental results show that it can achieve the separation of texture feature and the purpose of Feature recognition and detection.Finally,Kernel Fisher discriminant analysis(KFDA)method has demonstrated its success in extracting features.However,a single kernel is not always suitable for the different applications which contain data from multiple,heterogeneous sources.To improve the performance of KFDA,a novel algorithm named multiple data-dependent kernel fisher discriminant analysis is proposed in this paper.The constructed multiple data-dependent kernel is a combination of several base kernels with a data-dependent kernel constraint on their weights.By solving the optimization equation,the parameters optimization of data-dependent kernel and multiple base kernels are achieved.Experimental results on the three face databases validate the effectiveness of the proposed algorithm.更多还原