论文标题:数字图像和视频修复理论及其算法研究
Research on the Theory and Algorithms of Digital Image and Video Inpainting
论文作者
论文导师 马利庄,论文学位 博士,论文专业 Computer Application Technology
论文单位 上海交通大学,点击次数 755,论文页数 139页File Size5604K
2007-05-01论文网 http://www.lw23.com/lunwen_3947352/
Image inpainting;; Video inpainting;; Texture synthesis;; Temporal correlation;; Global motion estimation;; Optical flow analysis
数字修复理论与算法是近年来数字媒体领域倍受关注的研究热点。数字修复技术通过适当的算法估计并填充图像指定区域内的缺损数据。本文主要围绕静态图像以及动态视频图像序列内指定区域的修复问题展开研究。 1.在静态图像修复研究方面,本文概括了现有的图像修复理论与算法,对其中几种经典算法进行了详细的分析与比较,提出了一种新的基于纹理合成的图像修复算法。该算法不仅加快了修复速度,而且改善了算法的修复效果,提高了算法的鲁棒性,减少了误匹配及误差扩散的问题,尤其在针对较大尺度的区域进行修复时效果更为明显。其中: (1)提出了新的优先权计算方法。它通过引入全局梯度阈值,并对信心因子和数据因子规则化,使梯度较高的图像边缘得到优先修复,同时又保证了信心因子较高而梯度较弱区域的修复优先权。 (2)提出了新的样例图像块搜索策略。通过利用当前待修复图像块内的已知像素信息,预测修复后图像块的统计属性,筛选满足限制条件的待匹配的图像块计算其匹配代价。并通过引入双射因子,改进了匹配代价函数,解决了修复结果中部分图像细节重复出现造成的不自然问题。 2.在视频图像序列修复方面,本文提出了一种基于时间相关性的视频修复算法。其基本思想是利用时间邻域窗口内的图像信息修复当前帧内的未知像素。该算法优点是不限制摄像机的运动,不需要专门的视频前景-背景提取,便于对运动较为复杂的视频进行修复处理。 (1)研究了视频全局运动模型及其参数优化估计的方法。提出了在自适应的快速块匹配或Harris角点匹配的基础上,通过运动矢量直方图和最大类间方差两种方法剔除外点后,利用牛顿-拉夫逊方法进行迭代求解的全局运动估计方法。 (2)提出了在当前时间窗口内图像帧对齐操作的基础上,对当前帧中受损像素进行视频前景-背景标记,并分别修复的算法。该算法通过Poisson方程对未知像素进行预先填充,并在多分辨率Lucas-Kanade光流分析的基础上,对受损前景像素的局部运动矢量进行估计。与其它视频修复算法相比,该算法通过在当前时间窗口内进行色彩调整,使邻域图像帧与当前图像帧的色彩一致,从而解决了视频帧间差异造成的修复问题。
Research on the theory and algorithms of digital inpainting has attracted remarkable attention in recent years. The main goal of digital inpainting is to estimate and fill the pixels in some appointed image regions by proper method. This paper mainly studies on the inpainting problems of appointed regions in static images or video image sequences. 1. In the research on the static image inpainting, the existing theory and algorithms of image inpainting are introduced in generalities, and some representative inpainting methods are analyzed and compared in details. A new image inpainting algorithm is proposed based on texture synthesis. It accelerates the inpainting process, improves the inpainting results, enhances the robustness of the algorithm, and decreases the inpainting error and error diffusion effects, especially for large image regions inpainting. (1) A new priority computation function is proposed in our image inpainting algorithm. By introducing a global gradient threshold, the image edges with high gradients are inpainted preferentially, and at the same time the inpainting priorities of the image regions with high confidence terms and low gradients are also ensured by regularizing the confidence term and the data term. (2) A new strategy to search the example image patches is proposed in our image inpainting algorithm. From the known pixels in the current inpainting patch, the statistical property of the image patch after inpainting is forecasted and only the source image patches which meet some limitation are selected to compute the matching costs. Moreover, the matching cost function is improved by introducing a bijective-mapping term, which solves the artificial problem cause by some image details repetition in the final inpainting image. 2. In the research of video image sequence inpainting, a video inpainting algorithm based on temporal correlation is proposed in this paper. The main idea is to inpaint the unknown pixels in current frame by using the image information in the temporal neighbor frames of the video. The main advantage of new video inpainting algorithm is that it does not restrict the camera motion and can be used to inpaint the videos with complex motion without special video foreground-background extraction. (1) After the research on some video global motion models, a new video global motion estimation method is proposed based on adaptive rood pattern search or Harris corner matching, in which Newton-Raphson method is adopted to optimize the model parameters after removing outliers by the methods of motion vectors histogram and between-class variance maximization. (2) Based on frames alignments in current temporal neighbor, a new algorithm is proposed by tagging the unknown pixels with foreground or background and repairing them separately. In this algorithm, multiscale Lucas-Kanade optical flow analysis is adopted after repairing the unknown pixels by Poisson equation optimization to estimate the local motion vectors of unknown foreground pixels. Compared with other video inpainting algorithm, in this algorithm the inpainting problems caused by color difference between video frames are solved by adjusting the color of neighboring frames to make them consist with the current frame.
|
