面向感知的单幅图像分层深度信息计算

发布时间：2018-03-17 00:15

本文选题：图像分层　切入点：深度恢复　出处：《山东大学》2015年博士论文　论文类型：学位论文

【摘要】：赋予计算机认知和理解三维世界的能力是计算机视觉的一个基本问题,在机器人、虚拟现实与增强现实、智能监控、数字娱乐等领域有非常多的应用。目前数字图像的获取变得非常容易,互联网上存在着海量的图片,基于图像、特别是基于单幅图像恢复场景中蕴含的三维信息是目前研究的热点问题。由于物体之间的遮挡以及物体的自遮挡,基于单幅图像完全地恢复三维信息是一个病态问题,但人们对物体的认知有很多先验知识,有效地利用人们的先验知识使得从单幅图像中获取场景重要的三维信息变得可能。本学位论文选择图像分层深度信息计算这个基础问题作为研究的切入点,从人的感知特点出发,重点研究了基于单幅图像的两种分层深度信息计算方法,并将算法应用在两个实际问题中一浅浮雕艺术风格生成以及三维立体幻象的自动生成。除此以外,本文还研究了一种基于分层信息的自适应图像显著性引导算法,探讨了粗糙的层次深度信息在计算机视觉领域的其它可能应用。本文创新点和贡献主要包括以下几个方面：(1)提出了两种新的图像分层深度计算方法本文基于轮廓以及物体对象等两种不同的图像描述,提出了两种新的图像分层深度计算方法。本文利用轮廓线中的几何元素(包括区域、线段及结点)构建图(Graph)，根据有向图理论推导图像中所蕴含的深度层次。本论文还提出了一种基于物体对象的分层深度计算,根据单目视觉线索,构建马尔可夫图,从而估计图像中的深度层次,算法不仅能够有效推测局部层次关系,对于全局层次关系也有较高的估计准确度。(2)提出了一种基于单幅图像的浅浮雕生成算法本文通过计算图像中物体的轮廓、亮度、梯度等信息,基于层次深度计算,设计了一种基于反馈的浅浮雕逐步精细化算法,该算法不仅能够保持全局的层次关系,还能增强浅浮雕曲面的细节,从而生成自然生动的浅浮雕模型。(3)提出了一种基于单幅图像的三维立体显示技术本文提出了一种基于单幅图像分层深度信息的三维立体幻象自动生成方法。该方法根据物体的对称性及局部凸性等特点,利用粗糙的三维模型对图像中的对象进行表示,并对被遮挡部分自动进行三维补全,同时提出了一种基于迭代优化的地面拟合算法估计图像中物体的深度位置。生成的三维立体显示效果不仅能够使用诸如红蓝眼镜、偏振镜片等辅助设备观看,还可通过左右旋转视点提供裸眼三维体验。尽管该算法不能恢复精确的三维信息,但产生的三维立体幻觉效果符合人的感知特点,让人体验逼真的三维效果。(4)提出了一种自适应的图像显著性引导策略本文探索了图像分层深度计算与图像显著性引导之间的关系,提出了一种基于图像分层深度信息的自适应图像显著性引导算法。该算法分析图像中物体几何位置、深度排序等对于显著性的影响,并以其为特征指导对图像的编辑,在不对图像场景造成感知破坏的前提下增强物体显著性,从而引导用户视觉注意力。
[Abstract]:Give the computer the cognition and understanding of the three-dimensional world ability is a basic problem in computer vision, robotics, virtual reality and augmented reality, intelligent monitoring, digital entertainment etc. there are a lot of applications. At present, the digital image acquiring becomes very easy, on the Internet there are massive images based on the image, especially the single image recover the information contained in the scene is a hot issue of the current research. Based on the object between the object due to occlusion and self occlusion, a single image to fully recover 3D information is an ill posed problem based on the cognition of the object but there are a lot of prior knowledge, effective use of people's prior knowledge so as to get 3D information important from a single image becomes possible. This dissertation choose hierarchical image depth information to calculate the basic problems as the entry point, from Based on the characteristics of human perception, focus on the two calculation methods of layered depth information based on a single image, and applied the algorithm to generate a bas relief style in the two practical problems and the automatic generation of three-dimensional illusion. Besides, this paper also studies a guidance algorithm of adaptive image saliency layered information based on the study level of rough depth information in the field of computer vision applications. Other possible innovations and contributions include the following aspects: (1) put forward two new calculation method of the layered depth image contour and object in two different images based on the description, this paper proposes two new image layered depth calculation method. This paper uses the geometric elements in its outline (including regional, segments and nodes) to construct graph (Graph), according to the directed graph theory is contained in image The depth of the level. This paper also proposes a hierarchical object based on depth calculation, based on monocular visual cues, construct the Markov graph, to estimate the depth image in the algorithm can not only effectively infer local hierarchy, for global hierarchy has high estimation accuracy. (2) proposed a a shallow relief based on single image generation algorithm by calculating the image object contour, brightness gradient information, based on the calculation of depth, design a shallow relief based on the feedback of stepwise refinement algorithm, the proposed algorithm can not only maintain the hierarchy of global, but also enhance the bas relief surface details, so as to generate light natural and vivid relief model. (3) proposed a three-dimensional image display technology based on three-dimensional image information based on layered depth Vision automatic generation method. The method according to the object characteristics of symmetry and local convexity, the object in the image representation using three-dimensional model of rough, and the occluded part of 3D automatic completion, and proposes an iterative optimization based on surface fitting method to estimate the depth of 3D object in images. Stereo display generation can not only use such as red and blue glasses, polarized lenses and other auxiliary equipment to watch, but also through the rotation around the naked eye view provides 3D experience. Although this algorithm cannot recover the information accurately, but the three-dimensional illusion effect produced in accordance with the characteristics of human perception, let people experience a realistic three-dimensional effect. (4) proposed an adaptive image saliency guidance strategy this paper explores image layered depth calculation and significant relationship between image guide, propose a method based on Adaptive image saliency layered depth information guide algorithm. The algorithm analysis of the geometric position of the object in the image, the depth ordering of significant influence, and its guidance on image editing features, resulting in significant enhancement of object perception in the premise of failure not image scene can lead users visual attention.

【学位授予单位】：山东大学
【学位级别】：博士
【学位授予年份】：2015
【分类号】：TP391.41

【共引文献】