【作者】 张铁君；

【导师】 牛夏牧；

【作者基本信息】 哈尔滨工业大学 ， 信息安全， 2013， 博士

【摘要】 信息化深刻地改变了人类的生产与生活方式，在这一发展历程中，知识型经济与信息技术相生相成，受益良多。卡通作为最具有代表性的一类知识型经济产业在信息化过程中迎来空前的发展良机，其生产、传播、消费中的每一个环节都深受信息技术的影响，获得了迅猛发展。然而作为以版权为核心盈利模式的产业，卡通在信息社会中也面临着更为复杂的版权保护问题。卡通产业的版权问题主要在于卡通媒体在发布、传播、应用过程中未经授权的浏览、分享、仿制等行为，侵权行为呈现载体多样化、跨行业的特点。传统的基于数字签名或数字水印的版权保护技术已经难以应对针对卡通的复杂的侵权行为，产业界对面向卡通产业的版权保护技术的需求十分迫切。卡通角色是卡通作品的关键内容和衍生价值最高的部分，也是侵权行为最主要的目标；无论是对卡通作品的直接盗用还是对卡通作品的仿制，卡通角色都是其中不变的主题。所以卡通版权保护的核心就是对卡通角色的保护。本文以图像为对象研究卡通产业的版权问题，主要研究针对卡通角色的侵权行为的搜寻、过滤以及卡通媒体的机密性保护问题。首先探讨了卡通图像的基本性质，并指出形状是构成卡通图像的最基本语义内容，具有形状的相似性是卡通角色在不同性质的图像中内在联系。提出自然图像及卡通图像中线条的提取方法，来构成图像中的基本形状；随后提出了局部不变形状特征，以在位置、尺度、角度都未知的复杂图像中，对卡通角色的内容进行表述。进而展开对卡通角色的检索、检测以及选择加密研究，以解决卡通产业所面临的版权保护问题。本文的主要研究工作和创新点在于：(1)针对自然图像与卡通图像分别提出了检测其中边缘及线条的方法来提取图像中的形状信息。针对基于学习的边缘检测中存在的标记不准及噪声数据干扰的问题，提出了基于FSVR学习的边缘检测算法。该算法根据训练数据的可信度来确定训练数据的模糊权值，利用模糊权值来将低可信度数据对训练的影响降低，实验分析表明该算法可以获得较同类算法更佳的边缘检测效果。随后本文研究了卡通图像中线条的分类，分析了装饰线与边缘在一阶、二阶高斯导数滤波响应的不同特性，提出了新的针对卡通图像的线条提取算法，所提出算法较同类算法具有更佳的视觉效果与更清晰区分装饰线与边缘的能力。(2)提出了一种对尺度、方向变化都鲁棒的局部形状特征，并将该特征应用于对包含卡通角色形象的卡通图像、自然图像的检索中。该特征以卡通图像中常见的角点为特征点，对特征点附近邻域的边缘/线条的空间分布进行建模，从而获得对局部形状进行具有一定区分性与紧致性的描述。随后采用特征匹配的方法获得查询图像与待检索图像之间相似度。局部形状特征跨越了自然图像与非自然图像之间外观差异的鸿沟，同时排除背景信息的干扰，实现对局部内容进行检索。(3)针对卡通图像互联网分享过程中的滥用问题，提出了基于Hough voting的卡通角色检测方法，供媒体分享网站对用户上传图像或视频进行过滤。首先采用人工分割的模板来学习表达卡通角色的特征词典及基于Hough voting的目标表达。随后对模板和待检测图像特征进行匹配基础上，根据目标的表达，在Hough voting空间中计算某个位置存在该卡通角色的后验概率，从而获得该角色出现的位置及相应的可能性。算法在对卡通图像的检测中显示出了其准确性及效率。(4)提出了基于选择加密的卡通媒体分级预览方案，以在计算代价较低的情况下保证数据的机密性，并提供分级浏览的功能。首先基于薄板样条函数对已经定位于窗口内部的卡通角色精确提取外轮廓；随后基于混沌系统构造级联的序列密码算法对目标区域进行加密。选择加密的方法对核心内容进行了加密，而保留背景区域供人预览，提供了更好的用户体验性，提高了卡通媒体可推广性。更多还原

【Abstract】 Nowadays, information technology changes human being’s life a lot. In thisprogress, knowledge-based economy and information technology influence eachother, and make greate advances. Cartoon industry is a typical knowledge-basedindustry, and gains a good opportunity for further developing. All the aspects ofcartoon industry, including the producing, broadcasting and consuming, benefit fromthe information technology and make a great progress. However, the informationtechnology also brings more complex copyright problems at the same time, whichmake the cartoon copyright protection an urgent problem.The key issues of the copyright infringement problem in cartoon mainly lie onunauthorized browsing, sharing and copy during the cartoon broadcasting, spreadingand the application. Traditional copyright protection techniques, such aswatermarking and cryptography, are not sufficient for such complex problems.Obviously, cartoon characters are the key content of the cartoon media, which isalso the mostly frequently infringed part. Most of the infringements involve thecartoon characters, so the characters are the invariant theme in the infringements.Therefore, cartoon character protection is the key of the cartoon media copyrightprotection.This thesis focuses on the copyright protection of cartoon media. Firstly, thespecialty of the cartoon image is discussed, and a basic viewpoint is proposed: shapeis the essential element to present the cartoon character and the similarity in shape isthe internal connection between same characters in different images. Secondly, edgeand curve detection methods are proposed to extract the shape in either natural andcartoon images. Then, a new local invariant shape feature is proposed to describethe cartoon characters in the image, in which the location, angle and scale ofcharacter are unknown. Based on this feature, the research on the cartoon characterindexing, detection and selective encryption are done.The main contribution and innovation of this dissertation are described asfollows:Firstly, the learning based edge detection methods for natural images arediscussed and then a new edge detection method based on Fuzzy Weighed SupportVector Machine for Regression is proposed to alleviate the label impression problemand noise. A fuzzy membership is applied to re-weigh the training sample accordingto the distance to the human labels, so that the samples with different reliability aretreated differently to alleviate the label imprecision problem. Then a curveextraction method for cartoon images is proposed to extract the shape in the cartoon image. The two types of curve in the cartoon image are defined and extractedseparately based on their different response to the first and the second orderdirectional derivative. The proposed method is more efficient and produces betterresults.Secondly, a scale and rotation invariant local feature is proposed to describe thelocal shape, and used in the cartoon character retrieval in both the natural image andthe cartoon image. The feature locates at the scale invariant corner and use the shapecontext to describe the shape near the corner, and a feature matching method is usedto compute the similarity between the query and the image in the database. The usedof shape feature fill the gap between the appearance differences between naturalimage and cartoon image, and reduce the disturbance from the background.Thirdly, a Hough voting based cartoon character detection method is studied topreventing the unauthorized sharing in the website. The visual shape words areextracted in the pre-segmented cartoon characters and then a Hough voting schemeis used to present the character. And then a posterior probability conditioned on theexistence of visual words is estimated in the Hough voting space. The proposedmethod is effective as the experimental results show.Fourthly, selective encryption method is investigated in order to protect thecartoon from unauthorized copy. The Thin Plain Spline-Robust Point Matchingmethod is employed to localize the cartoon character accurately, and then a streamcipher based on multi-chaotic systems is used to encrypt and decrypt the characterregion. The selective encryption method encrypts the key content of the cartoon andleaves the others as plain for the preview, which can offer better user experiencewhich make cartoon publish more easily.更多还原