全景视频图像融合与拼接算法研究

发布时间：2018-05-19 01:10

本文选题：图像拼接 + 虚拟现实　；参考：《电子科技大学》2017年硕士论文

【摘要】：虚拟现实(Virtual Reality,VR)技术普及和市场拓展在最近五年有一次大的飞跃。特别是VR 360全景视频的商业应用已经走入千家万户,进入了快速发展时期。人们对主观质量、沉浸式感受的需求越来越高导致高分辨率、高动态范围的全景视频数据量骤升。这直接引发了VR视频数据在传输和储存两个方面同时面临更大挑战。因此,研究VR全景视频合成和编码压缩具有强烈的现实急迫性。论文以全景视频图像的融合、拼接、编码相关技术为主要研究内容,具体地,在四个方面展开工作。(1)实现全景视频图像采集、预处理、配准到融合的全过程。基于正六面体六镜头摄像机采集的视频数据,采用棋盘标定法对相机进行标定,采用白平衡算法解决图像色差一致性问题。将六幅图像映射到球面,采用SIFT算法进行配准,采用重叠区域线性过渡法进行融合,最后生成VR 360全景视频图像。(2)分析VR球面图到平面图的多种映射格式,选择最佳全景视频编码输入格式。分析对比圆柱体、六面体、八面体、二十面体、新型瓦片分割等映射转换方式在编码性能、主客观质量及计算量方面的差异。验证得出六面体是目前最佳的编码输入格式的结论。(3)基于图像融合拼接技术提出全景视频运动估值越界折叠搜索算法。利用全景视频图像左右边界的连续性,运动补偿时通过图像拼接技术将重构帧左右边界进行拼接融合,降低运动估值搜索匹配块的误差,提升了编码效率。(4)针对CIS扫描仪扫描图像去噪提出了一种光栅条纹去除算法。分析了CIS扫描仪物理结构,建立了光栅条纹噪声数学模型,提出了傅里叶函数算法,有效去除了图像的条纹噪声。上述四个方面的工作有望为全景视频融合、拼接、编码的算法研究提供一些有益参考。在全景图像融合与拼接算法方面达到了合成全景图像平滑自然,具有较好的主观质量。对多种VR球面图到平面图映射格式测试结果表明,六面体为现有的优于经纬图的映射格式,建议作为全景视频编码输入格式的主要选择。全景视频运动估值折叠算法能够找到失真更小的匹配块,有效提升了编码效率。光栅条纹去除算法能有效去除CIS扫描图像噪声,且比现有噪声去除算法效果更佳。
[Abstract]:Virtual reality (VR) technology popularization and market expansion have made a big leap in the last five years. In particular, the commercial application of VR 360 panoramic video has entered a rapid development period. The increasing demand for subjective quality and immersive experience leads to a surge of panoramic video data with high resolution and high dynamic range. This directly leads to greater challenges for VR video data transmission and storage at the same time. Therefore, it is urgent to study VR panoramic video synthesis and coding compression. In this paper, the fusion, splicing and coding technologies of panoramic video images are taken as the main research contents. Specifically, the whole process of panoramic video image acquisition, preprocessing, registration and fusion is realized in four aspects. Based on the video data collected by the hexahedron six-shot camera, the camera is calibrated by the checkerboard calibration method, and the consistency of the image color difference is solved by the white balance algorithm. The six images are mapped to the sphere, the SIFT algorithm is used for registration, and the overlapping region linear transition method is used to fuse the image. Finally, the VR 360 panoramic video image. 2) is generated to analyze the various mapping formats from the VR spherical image to the plane map. Select the best panoramic video encoding input format. The differences of coding performance, subjective and objective quality and computational complexity between cylindrical, hexahedron, octahedron, icosahedron and new tile segmentation are analyzed and compared. The conclusion is that hexahedron is the best coding input format at present. Based on image fusion and stitching technology, the paper proposes an algorithm for searching panoramic video motion estimation by overstepping folding. Using the continuity of the left and right boundaries of panoramic video images, the reconstructed frame left and right boundaries are stitched and fused in motion compensation, which reduces the error of motion estimation search matching block. A raster stripe removal algorithm is proposed for CIS scanner image denoising. The physical structure of CIS scanner is analyzed, the mathematical model of grating fringe noise is established, and the Fourier function algorithm is proposed to remove the fringe noise effectively. These four aspects are expected to provide some useful references for the research of panoramic video fusion, mosaic and coding algorithms. In the aspect of panoramic image fusion and stitching algorithm, the synthetic panoramic image is smooth and natural, and has good subjective quality. The test results show that the hexahedron is a better mapping format than longitude and weft map. It is suggested that it is the main choice of input format for panoramic video coding. Panoramic video motion estimation folding algorithm can find matching blocks with less distortion, which effectively improves the coding efficiency. Raster stripe removal algorithm can effectively remove the noise of CIS scanning image, and the effect is better than the existing noise removal algorithm.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】