面向MPEG-4的视频对象分割算法研究

【作者】 罗玲；

【导师】 解梅；

【作者基本信息】 电子科技大学 ， 信号与信息处理， 2003， 硕士

【摘要】 随着通信和信息处理技术的发展，基于视频的应用呈现出强大的灵活性和可扩展性，视频通讯已经成为增长最快的信息载体，大量数字化的视频服务需求不断涌现，如：数字电视、远程会议、视频电话和交互式多媒体等。这些伴随着大数据量的应用和服务需要更先进的数字信号处理技术，以便进行更高效的存储和传输，以及更准确的分析和更灵活的操作。 MPEG-4标准正是为解决这些问题应运而生的。MPEG-4最重要的特点在于用音频—视频对象来描述内容和进行编码。在MPEG-4标准中，视频帧被认为是由一系列相互独立的运动对象组成，视频帧的编码也直接针对一个个视频对象。所以，视频对象的分割提取是MPEG-4标准的核心问题，这是一个非常复杂的问题，在MPEG-4标准中没有具体的算法规定，到目前为止还没有一个通用的方法。因此，本文选择了基于MPEG-4的视频对象分割算法这一课题研究，具有重大的理论和实用价值。 本文研究视频对象分割方法，采用了对空域进行分水岭变换，对时域进行运动检测，再结合时空信息共同判定视频对象的方法。侧重于视频对象分割的准确性、通用性，同时尽量减少算法的复杂度和计算量，以便能够适应于视频通信的实时环境。本文章节安排如下：第一章描述蓬勃发展的多媒体技术、MPEG标准及新的MPEG-4标准对视频对象分割提出的新要求。第二章介绍现有的视频对象分割技术。在第三章中，介绍了分水岭变换的定义、优缺点，并研究了在传统分水岭的基础上利用数学形态学滤波器缓解算法中过分割现象的改进算法，并给出一种能获取全局最优解的区域合并方法。第四章是对运动信息检测技术的介绍及实验结果讨论，这章涉及到时域上的运动掩码检测方案，全局运动补偿，基于概率模型的运动检测以及后处理。第五章介绍时空联合提取技术，在这章中我们研究了采用双阀值进行前、后景判定的方法，提高了提取对象的准确性。最后一章是对整个算法的总结、回顾和对其前景的展望。在文中每章节中我们都给出了相应实验结果，用于验证该章节中算法的合理性。更多还原

【Abstract】 With advances in communication and information processing technologies, the video-driven application show a very large degree of flexibility and extensibility. Visual communication is the fastest growing vehicle for information. A lot of digital applications and services are emerging, such as: digital TV, teleconference, videophone and interactive multimedia. These diversified applications and services with a large amount of data demand more advanced digital signal processing technologies for efficient memory and transmission, accurate analysis and flexible manipulation.MPEG-4 standard is such a standard to solve above problems. Audio-video object-based coding and description are the distinct features of MPEG-4 standard. In the MPEG-4 standard, video frame was composed of a series of independent semantic video object, which was encoded independently. So, the performance of the segmentation algorithm is crucial to the final MPEG-4 coding products. Video segmentation is a challenging topic of research, no single algorithm has been reported in the literature that is generally applicable.In the dissertation we proposed a spatial and temporal VOPs’ Segmentation and Extraction methods for MPEG-4. I concentrate my research work on accurate , common and automatic algorithms, At the same time reduce the complexity and computation burden of algorithm as could as possible to match the needs of real-time requirement. This paper was organized as follows: First, an overview about the multimedia technology, MPEG standards and some research backgrounds were shown. Second, some popular spatial and temporal technology about the VOP segmentation were introduced. Third, a modified watershed segmentation preprocessed by multi-scale morphological filters and post processed with improved fast region merging methods was proposed to obtain a good spatial segmentation . Forth, to extract the semantic object from the spatial segmentation, we applied a temporal segmentation based on the global motion compensation, statistical model-based change detection and post processing technology to get the motion mask. Fifth, a methods based on the above spatial and temporal result was brought forward to obtain the final, a two threshold method was proposed. Finally, we summary our algorithm and propose the improve direction in the future .At each stage of our algorithms, some experimental result was given to validate the performance of the proposed algorithms.更多还原