高准确度深度获取与应用研究

发布时间：2018-04-11 01:23

本文选题：多视点图像 + 深度图像　；参考：《上海师范大学》2017年硕士论文

【摘要】：与传统的2D和3D电视相比,多视点能允许用户在一定的范围内自由选择视角,观看其感兴趣的内容,因此可以给用户带来更震撼的视觉冲击力、丰富的立体感和非比寻常的三维沉浸感。相比于当前的高清视频,多视点视频的数据量会随着视点数的增加呈线性增长,目前的存储介质和网络带宽显然无法满足多视点立体视频的要求。而基于深度图像的绘制(DIBR,Depth Image Based Rendering)技术是多视点视频的关键技术之一,它有效的结合了纹理图像和对应的深度图像的信息,使得任意新视点图像中的各个像素点都可以通过3D图像映射方程得到。该项技术凭借其节省数据空间和网络带宽、质量高、绘制速度快的优点而被公认为是下一代三维立体视频的候选方案。然而,获取高质量的深度图是十分具有挑战的,它的准确度直接影响立体视频的绘制质量。针对上面的问题,本文的主要工作和成果如下:1.提出了一种基于Kinect的高准确度深度获取的图像增强算法,从Kinect获取背景纹理图,并采用形态学滤波与双边滤波相结合的方法填充空洞。采用三维高斯混合模型(3DGMM,3D Gauss mixture model)来分离深度图像的前景和背景。针对空洞出现的区域做不同的处理。背景空洞采用对应的背景值的填充,前景空洞采用颜色匹配和结构相似性结合的算法填充,获取到可靠的深度图像。2.提出了一种改进的基于方向纹理的虚拟视点合成方法,将配准后的纹理图和最终的深度图经三维映射后得到虚拟视点的纹理图和深度图,映射后的新视点图像中的空洞存在的区域分为前景空洞和背景空洞,背景区域的空洞采用基于背景值的填充,前景区域空洞采用改进的基于方向纹理的虚拟视点合成的图像修复技术进行修复,得到较好虚拟视点图像。3.提出了一种基于感兴趣区域的多视点颜色校正算法并进行质量评价。首先获取多视点视频的源图像和目标图像,然后提取目标图像的感兴趣区域的显著性图,对显著图和源图像对应区域进行匹配,最后对匹配后的图像进行主客观质量评价。实验结果表明,与传统的三维高斯混合模型的算法相比,该算法的客观质量评价结果与主观感知质量有较好的一致性。
[Abstract]:Compared with traditional 2D and 3D TVs, multi-view allows users to freely select their views and view the content of their interest within a certain range, thus giving users a more shocking visual impact.Rich three-dimensional and unusual three-dimensional immersion.Compared with the current high-definition video, the amount of multi-view video data will increase linearly with the increase of the number of views. The current storage media and network bandwidth obviously can not meet the requirements of multi-view stereo video.The depth image rendering Image Based rendering is one of the key technologies of multi-view video, which effectively combines the texture image and the corresponding depth image information.Each pixel in any new view image can be obtained by 3D image mapping equation.Because of its advantages of saving data space and network bandwidth, high quality and high rendering speed, this technology is recognized as a candidate for the next generation 3D stereo video.However, obtaining high quality depth maps is a challenge, and its accuracy directly affects the quality of stereo video rendering.In view of the above questions, the main work and results of this paper are as follows: 1.An image enhancement algorithm with high accuracy and depth acquisition based on Kinect is proposed. The background texture image is obtained from Kinect, and the cavity is filled with morphological filtering and bilateral filtering.The 3D Gauss mixture model of 3D Gao Si is used to separate the foreground and background of depth image.Do different treatment for the area where the void occurs.The background cavity is filled with the corresponding background value, and the foreground cavity is filled with a combination of color matching and structural similarity to obtain a reliable depth image.An improved virtual viewpoint synthesis method based on directional texture is proposed. The texture map and depth map of virtual view are obtained by 3D mapping of the registered texture image and the final depth map.The voids in the mapped new view images can be divided into foreground voids and background voids, and the voids in the background regions are filled with background values.The voids in the foreground region are repaired by an improved image restoration technique based on directional texture-based virtual view synthesis, and a better virtual view image .3is obtained.A multi-view color correction algorithm based on region of interest is proposed and its quality is evaluated.Firstly, the source image and target image of multi-view video are obtained, then the salient map of the region of interest of the target image is extracted, and the corresponding regions of salient image and source image are matched. Finally, the subjective and objective quality evaluation of the matched image is carried out.The experimental results show that the objective quality evaluation results of this algorithm are in good agreement with the subjective perceived quality compared with the traditional three dimensional Gao Si hybrid model.
【学位授予单位】：上海师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TN948.6

【参考文献】