【作者】 许文丹；

【导师】 来新泉；

【作者基本信息】 西安电子科技大学 ， 电路与系统， 2014， 博士

【摘要】 随着视频技术的迅速发展及广泛应用，提供快速、有效及自动化的图像序列表达及处理方法已经成为了一个重要的研究领域，其中基于图像内容的方法，例如基于目标及特征区域的方法已经成为许多应用的首选。由于此类方法能够有效地消除图像序列中的“握手效应”，快速提取图像序列中的目标，以低比特率传输图像的形状、运动及纹理信息，并保持图像稳定性，因此它们在现代互联网、广播、电视、娱乐及第四代移动通讯等领域正被广泛采用。本论文介绍了作者对低比特率图像序列编码算法及框架的研究，重点研究了基于内容的图像目标分段、稳定性算法及实现框架设计。本文的目的是研究如何寻找简便高效的算法并提出一个集成这些算法的实现框架，使理论与实验成果走向实际应用。作者的主要研究成果包括：首先，在研究以往的目标分段算法的基础上，提出了一种用于目标识别的自适应变化检测新算法。该算法采用一个三步法，能快速有效地把图像目标从背景中分离出来。第一步是依据亮度差及照度变化，分别把图像序列中的噪声和运动目标识别并分离出来；第二步是利用图像块、直方图及区域分类，把图像分割成为与运动目标相对应的区域；第三步是在前两步的基础上，进行形态边缘检测、轮廓分析及目标标识，以完成最终的目标识别、图像分段任务。其次，作者在上述图像目标识别、分割算法的基础上，设计了一个新的低比特率图像序列编码方案，该方案利用图像变化区域内的运动矢量信息、图像形状角点信息及无运动或准静止区域内的余留信息来完成高效视频压缩，其编解码性能优于传统的典型的编码算法。此外，本文针对实际应用中的图像序列不稳定现象（通常来自于摄像源），提出了一种新颖的图像运动补偿方法，该方法对来自于图像序列源的运动进行估计，并以补偿平移和旋转的方式来抵消此类运动。实验结果显示，该算法可以有效地稳定实时捕获的各类视频。为了验证本文提出和改进的有关算法，作者进行了大量的计算机模拟及实验，并同以往的传统经典方法进行了比较，说明了本文提出的若干方法取得了良好的效果。大量不同类型的实际图像序列实验也表明，本文提出的算法和方案性能可靠并优于文献中的典型算法，具有较好的应用前景。更多还原

【Abstract】 As the use of video becomes increasingly popular and wide spread in the areas ofbroadcast services, internet, entertainment and security-related applications, providingmeans for fast, automated, and effective techniques to represent video based on itscontent, such as objects and meanings, is important topic of research. In manyapplications, removing the hand shaking effect and making video images stable andclear or decomposing (and then transmitting) the video content into a collection ofmeaningful objects is a necessity. Therefore automatic techniques for video stabilization,extraction of objects from video data as well as transmitting their shapes, motion andtexture at very low bit rates over error networks, are desired.In this thesis the design of a new low bit rate codec is presented. Furthermore amethod for video stabilization is introduced. The main technical contributions resultedfrom this work are as follows.Firstly, an adaptive change detection algorithm identifies the objects from thebackground using a three-stage method. In the first stage, the luminance differencebetween framers is modelled in order to separate noise and illumination variations frommeaningful moving objects. In the second stage the segmentation tool based on imageblocks, histograms and clustering algorithms segments the difference image into areascorresponding to objects. In the third stage morphological edge detection, contouranalysis, and object labelling are the main tasks of the proposed segmentationalgorithm.Secondly, a new low bit rate codec is designed and analysed based on the proposedsegmentation tool. The estimated motion vectors inside the change detection mask, thecorner points of the shapes as well as the residual information inside the motion failureregions are transmitted to the decoder using different coding techniques, thus achievingefficient compression.Thirdly, a novel approach of estimating and removing unwanted video motion,which does not require accelerators or gyros, is presented. The algorithm estimates thecamera motion from the incoming video stream and compensates for unwantedtranslation and rotation.A synchronization unit supervises and generates the stabilized video sequence. Thereliability of all the proposed algorithms is demonstrated by extensive experimentationon various video test sequences.更多还原