【作者】 张青；

【导师】 肖春霞；

【作者基本信息】 武汉大学 ， 计算机应用技术， 2017， 博士

【摘要】 由于包含信息量丰富,且直观易懂,图像视频在数字传媒、智能系统、社交娱乐、网络直播、监控安防和军事侦查等多个领域都有重要应用。上述应用通常以对图像视频内容进行准确理解和可靠编辑为前提,而人类视觉感知系统对色调高度敏感,因此以色调为线索或载体进行图像视频内容理解和编辑受到了大量的关注,现已发展成为计算机图形学、计算机视觉和图像视频处理领域的一个热点研究问题。色调敏感的图像视频内容理解与编辑以色调信息为切入点,通过构建模型或基于相关理论,对图像视频中的内容进行可靠的分析、识别、理解和编辑,以生成满足实际应用需求的结果为目的。本文对该课题所涉及的理论、方法、关键技术以及应用上展开深入、系统的研究,围绕航空图像云检测、欠曝光视频增强以及图像重着色三个方面开展图像视频内容理解与编辑研究工作。具体地,本文首先对航空图像进行色调驱动的内容理解,探讨了航空图像云检测这一研究问题。接下来,重点研究了以色调编辑为主体的欠曝光视频增强和图像重着色。对图像视频进行快速、可靠的内容理解和编辑是涉及诸多困难和挑战的研究课题。对内容的理解要求准确区分场景中不同对象,并能推断出该对象是何物体、位于何处,其中牵涉到场景语义分析、目标识别和图像分割等多个问题。尽管上述问题对应的研究领域取得了长足的进步,但对于特定对象的全自动、高效率、高精度理解仍然是有待解决的问题。对于内容编辑的研究工作已经有很多,涉及颜色、光照、对比度、风格和时空位置等多个方面,但普遍存在效率低、可用性差和视觉感知不友好等问题。认识到人类视觉感知系统对于色调的高度敏感性,本文以色调为线索,多角度研究了图像视频内容的理解和编辑问题,为包括航空图像云检测、欠曝光视频增强和图像重着色三个领域的研究提供了行之有效的解决方案。本文首先从云区和非云区之间的色调差异性出发,深入研究了高分辨率RGB彩色航空图像云检测问题。接下来,对低照度环境下色调的恢复问题进行了分析和思考,探讨了欠曝光视频增强问题。最后,本文研究基于简易颜色面板如何对图像进行快捷有效的重着色。具体地,本文主要研究内容包括如下几个方面:(1)提出了基于累进式优化的RGB彩色航空图像云检测方法。方法以航空图像云区色调和纹理的统计观察特性为基础,构建了一种由粗到精、逐步细化的累进式云检测框架,有效区分了云区和非云区地物,获得了高精度云检测结果。另外,本方法首次引入了局部线性模型,用于检测航空图像中的半透明云区。(2)提出了基于感知驱动融合的欠曝光视频增强方法,达到了对欠曝光视频进行虚拟光照补偿,进而提升场景能见度、细节辨识度和颜色饱和度的目的。其中,本方法首次引人类视觉感知度量用于评估视频质量,引导算法生成全局最佳虚拟曝光的视频结果。另外,提出了感知驱动递进式融合框架,用于无缝融合所有局部最佳曝光的区域。最后,为消除视频中存在的噪声干扰,且保留清晰的纹理结构,提出了保纹理自适应时空滤波器。(3)提出了基于颜色分解优化的面板驱动图像重着色算法,显著简化了图像重着色这一常见图像编辑操作,同时保证了生成的结果符合人类视觉认知,具备逼真的颜色和自然的外观。为快速生成颜色面板描述图像中的主要颜色分布,本方法首先提出了一种与图像大小无关的高效面板提取算法。之后,构建了颜色分解优化模型,揭示了颜色面板和图像颜色之间的本质联系。最后,介绍了一种图像重着色算法评估思路,并首次公布了图像重着色测试图片集用于算法评估和对比。围绕着图像视频内容理解和编辑这一主题,本文多角度、全方位地对色调相关的问题进行了深入系统的研究,对其中涉及的关键技术难点提供了详细的解决思路和方法。实验结果表明,本文提出的算法在各自领域都优于已有的方法,且可以被推广到多种图像视频相关的应用场景。更多还原

【Abstract】 Due to the information richness and intuitive understandability,image and video have lots of applications in digital media,intelligent systems,social entertainment,webcasting,surveillance security and military investigation and other relevant fields.These aforementioned applications basically rely on accurate understanding and reliable editing to image and video.As the human visual perception system is highly sensitive to the tone,tone-aware image and video understanding and editing have attracted a lot of attention,and have been developed into a hot research topic in both computer graphics,computer vision and image and video processing.The main purpose of the topic is to help the computer or user fulfill tasks such as analysis,identification,cognition,adjustment and editing of image and video,and to generate the results that meet requirements of specific application through construction of suitable models and theories based on tone.In this paper,the theory,method,key technology and application of this topic are deeply studied.Specifically,the research is carried out in three aspects:cloud detection of aerial image,underexposed video enhancement and image recoloring.In particular,we first study the tone-driven content understanding problem,and explore the cloud detection of color aerial photographs.Next,we focus on the underexposed video enhancement and image recoloring with tone editing as the main body.Fast and reliable understanding and editing of image and video is a research topic that involves many difficulties and challenges.The procedure requires accurate dif-ferentiating of objects in the scene and inferring their categories and location,which essentially involves the scene semantic analysis,object recognition,image segmenta-tion and many other problems.Although researchers have made great progress in the field of these problems,efficient understanding of specific object is still a problem to be solved.There are many image and video editing approaches,involving color,illumina-tion,contrast,style and spatial-temporal location,etc.However,they commonly have low efficiency and poor usability,and fail to generate visually satisfactory results.Due to the high sensitivity of the human visual perception to tone,we in this paper use tone as clue,comprehensively study how to understand and edit image and video in three aspects,namely cloud detection of aerial photographs,underexposed video enhancement and image recoloring.Based on the tone difference between cloud and non-cloud region,we first study cloud detection of high resolution RGB color aerial photographs.Then,we analyze and discuss tone recovery in low light level environment,and explore the underexposed video enhancement problem.Finally,we study how to achieve intuitive,efficient,and realistic image recoloring by editing a simple color palette derived from the original image.Specifically,this paper has the following three main contributions:(1)we propose a cloud detection method for RGB color aerial photographs.Based on statistical characteristics of hue and texture of cloud region in aerial photographs,a coarse-to-fine cloud detection framework is constructed,which achieve high de-tection accuracy by effectively separating cloud and non-cloud ground regions.In addition,we introduce a local linear model for the first time to detect semitrans-parent cloud regions in aerial photographs.(2)We develop an underexposed video enhancement method based on perceptual-driven fusion,which achieves the purpose of virtual illumination compensation for underexposed video,and accordingly enhance scene visibility,contrast and color saturation.The proposed method makes the first attempt to incorporate the hu-man visual perception measures for assessing video quality,which helps guide the algorithm to generate the global well-exposed video.Besides,a perception-driven progressive fusion framework is proposed for seamless integration of all locally best exposed regions.Finally,in order to remove the noise interference and to avoid degrading texture structure,we further propose texture-preserving adaptive spatio-temporal filtering.(3)We propose a palette-driven.image recoloring algorithm based on color decompo-sition optimization,which greatly simplifies the image recoloring operation while ensuring that the generated results are consistent with human visual perception,with vivid colors and natural appearance.To efficiently summarize the main color categories of the original image into a compact color palette,we first proposed an effective palette extraction algorithm,which is independent of the image size.Then,we construct the color decomposition optimization framework for revealing thees-sential relationships between color palette and pixel colors.Finally,we introduce how to evaluate image recoloring algorithms,and firstly publish a image collection for evaluation and comparison.Focusing on understanding and editing image and video,we study several tone-related problems from different perspectives,and provides detailed solutions to key technical difficulties involved in each problem.The experimental results show that the proposed methods in this paper are superior to the existing approaches in their respective fields,and can be extended to a variety of image and video related application scenarios.更多还原