【作者】 周磊；

【导师】 周源华；

【作者基本信息】 上海交通大学 ， 通信与信息系统， 2008， 博士

【摘要】 旧电影修复技术是计算机视觉、数字图像处理、数字视频处理、数字节目非线性编辑等许多研究领域的核心技术和热点方向。本文针对胶片电影中存在的几种主要失真做了详细研究,其中主要研究了数字电影修复策略、数字修复过程中的基础技术、以及几种典型失真的数学模型,并在充分汲取现有视频处理理论及相关应用中技术优点的基础上,对稳像,划痕,闪烁,图像噪声和低亮度CCD噪声的检测和修复做了细致分析,针对它们在空时域上的数学特征,修复算法的计算复杂度以及图像恢复质量的不足进行了拓展性研究,提出了一些相对应的视频失真检测或恢复方法,从而在有效地增强了图像序列修复质量的同时,还提升了视频编码器处理该类电影序列时的性能。具体的,本论文的主要研究成果如下:首先,阐述了数字实时稳像技术的理论基础,回顾了两种典型的传统稳像算法,设计了一种实时视频稳像系统,包含完整的稳像处理步骤.其中全局运动分析,局部运动块剔除,切边时图像尺寸约束,帧扩边和稳像帧生成步骤都有一定的创新性设计。该系统具有三个工作模式,可以设置不同的输入输出视频尺寸,其对平移和旋转抖动的稳像效果优良,并且速度性能完全能适应低速嵌入式平台环境。其次,阐述了直线划痕的视觉特征,改进了划痕数学模型表达式,同时改进了空间直线划痕检测器的设计,并结合划痕在时间域上低幅度,与全局运动不相关的运动特征,设计了一种基于上述准则进行时域判决的补充算法。这种空时检测器算法具有更大的适用范围和更强的识别能力。第三,阐述了亮度闪烁的视觉特征,描述了多种闪烁数学模型,基于其中的不同模型回顾了SLAR修复算法和Roosmalen的迭代算法,分析了其优缺点,提出一种改进型的SLAR算法(ISLAR),以实现低复杂度快速修复,最后设计了一种基于块的多帧统计恢复算法。以达到更精确的修复效果。最后,提出了一种新的双向细节空时降噪滤波器设计:它由一个噪声估计器,一个运动检测器,一个双向细节检测器和一个三帧高斯自适应加权滤波器构成,使用前两个修复帧和当前帧来达到实时去噪的目的。该滤波器适合处理高斯噪声,面对脉冲噪声和混合噪声时可以增加一个脉冲噪点检测器和一个改进的中心加权中值滤波器来处理。此外,还提出一种针对CCD低亮度噪声的空时多帧平均降噪滤波器算法,其关键是带有一个严格的基于块的运动检测器,为了处理镜头摇摆带来的误判,还增加了全局运动估计和基于块的运动补偿。面对低亮CCD视频,该算法在高效修复的基础上,保持了较低的算法复杂度,有效地去除了低亮场景下的灰度噪声和彩色噪声,并且提升了编码器的编码帧率,降低了压缩流码率。更多还原

【Abstract】 Old film restoration is one of key technology and has drawn tremendous attention in research of computer vision, digital image processing, digital video processing and digital program non-linear editing technology. In this thesis, the digital film repair scheme, the basic theory of video restoration and mathematical models of several typical artifacts are investigated. By fully taking advantages of the existed video process theory technology on related applications, this thesis deeply analyzes the detection and removal method for stabilization, scratch, flicker, image noise and low light level CCD noise, extend the research upon above artifact features in spacial and temporal field, reduce the complexity of restoration algorithms and improve their repairing effect. In this thesis, we provide some novel corresponding detection and restoration algorithms, which not even enhance the quality of restored sequence, but also improve the efficiency of video encoder when compressing the film sequence degraded by these kinds of artifacts. The creativities and contribution are discussed in detail as follows:Firstly, the technological framework and basic theory of digital video real-time stabilization are presented. Here we review two typical traditional algorithms for video anti-shaking, then propose a real-time video stabilization system, which includes all steps for stabilized video generating. Some innovative designs are presented in global motion analysis, local motion block removal, constraint of limited edge for cropping, frame’s edge extending, stabilized frame reconstruction, etc. This system has three working modes, can support different sizes for input and output video. It not only gets good effect on smoothing translation and rotation unsteadiness, but also reaches a high speed performance and fits for many applications on low speed environment, such as the embedded platform.Secondly, we describe the visual feature of line scratch, improve the mathematical model and spacial method of line scratch detection, and then, we reinforce the precision of detection results with a novel temporal refining detector according to the especial motion feature of scratch, which is small and different from the global frame motionThirdly, we present the visual feature of intensity flicker, list several degradation models with different precision, and then we describe two previous recovery method: SLAR algorithm and Roosmalen’s iterative algorithm, analyzes the advantage and shortcoming of thesis methods. In order to realize a fast and low complexity restoration, we provide an improved SLAR method (ISLAR) to deal with the quickly moving object. On the other hand, for the sake of getting more accurate restored sequence, we design a block-based method based on a multi-frame statistical average algorithm, which is complicated but effective.Finally, a novel bidirectional detail-preserved spatio-temporal de-noise algorithm is proposed, which consists a noise estimator, a motion detector, a bidirectional detail detector, and a three-frame adaptive weighted filter. This method uses two previous restored frames and the current frame to reduce noise. The filter is suitable for Gaussian noise, but if it is employed to filter the impulse noise or mixed noise, an optional impulse noise detector and an optional CWM filter will be activated to work together. Furthermore, this thesis proposed a spatio-temporal multi-frame average filter to deal with low light level CCD noise. The key part of this filter is a strict block-based motion detector. In order to deal with the false judgment comes from the camera shaking, we employee a global motion estimation (GME) and block-based motion compensation. For the low light level CCD video, this algorithm takes a small amount of calculation but lead a high quality recovery. Actually the filter can not only remove most of luminance noise and color noise caused by LLL background, but also improve the encoding frame rate and reduce the data size of compressed bit stream efficiently.更多还原