多尺度变换域图像的感知与识别:进展和展望  

Advances and Perspective on Image Perception and Recognition in Multiscale Transform Domains

作　　者：焦李成[1] 孙强[2] 

机构地区：[1]西安电子科技大学智能信息处理研究所,西安710071 [2]西安电子科技大学雷达信号处理国家重点实验室,西安710071

出　　处：《计算机学报》2006年第2期177-193,共17页Chinese Journal of Computers

基　　金：国家自然科学基金(60133010;60372050);国家"八六三"高技术研究发展计划项目基金(2002AA135080);国家"九七三"重点基础研究发展规划项目基金(2001CB309403)资助~~

摘　　要：多尺度变换域隐马尔可夫模型能够有效地描述变换域系数在尺度间、尺度内和方向间的统计相关性,是一种新的统计图像感知与识别方法.文中以变换域系数的统计相关性描述为中心,以模型的设计和应用的开展为两翼,深入分析了子波变换的三级统计特性与机理,比较研究了多尺度变换域的十种统计模型,并系统评述了这些模型在图像感知、处理和分析中的最新进展.同时,具体论述了这一领域研究中两类成功的实例:图像去噪和图像纹理分割.对于前者,以Lena图像为测试用例分析比较了以变换域统计模型为核心的8种算法的去噪性能;对于后者,按照分割类型(监督式或非监督式)和应用的图像类型系统比较了以统计模型为基础建立的15种图像分割方法.最后,从面向应用的模型构造和算法设计、变换域的拓展和应用层次的推广三个层面指出了目前存在的问题和不足,探讨了进一步的研究重点.Muhiscale transform-domain hidden Markov models（HMM＇s）, a class of new statisti cal approaches to image perception and recognition, can effectively characterize the interscale, intrascale and cross-orientation correlations of the coefficients in different multiscale transform domains. With the statistical characterization of the coefficients in transform domains the center and the design of various statistical models and the development of a variety of applications two wings, three classes of statistical characteristics and their mechanisms of discrete wavelet transform of an image are analyzed in depth, ten statistical models in multiscale transform domains are studied comparatively, and the state-of- theart of multiscale transform-domain statistical models in image perception, processing and analysis is systematically reviewed. Two types of successful applications in this research domain are discussed in detail, i. e. , image denoising and textured image segmentation. With respect to the former, with Lena image as a test case, the denoising performances of different algorithms based on eight kinds of transform-domain HMM＇s are com pared in terms of peak of signal-to-noise ratio（PSNR）; for the latter, fifteen methods derived from multiscale transform-domain HMM＇s are systematically compared according to the segmentation categories（supervised or unsupervised）and the image types on which these methods are applied. Finally, several main problems and a few deficiencies are pointed out and further challenges foreseen from three aspects, the application-oriented model construction and algorithm design, the generalization of multiscale transform domains, and the extension of application domains.

关 键 词：子波分析 隐马尔可夫模型 统计建模 图像感知 多尺度几何分析 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]