压缩感知光场重建及深度估计的研究

发布时间：2018-05-04 05:08

本文选题：光场 + 压缩感知　；参考：《郑州大学》2017年硕士论文

【摘要】：光场图像包含丰富的空间3D信息,因此光场图像可以用于重聚焦、深度估计以及三维显示,其中精确的深度信息对显著性检测、超分辨重建、目标识别及3D表面重建等技术的发展具有重要的作用。虽然相机阵列和微透镜阵列光场图像获取方法能够有效的记录光场图像,但相机阵列方法由于体积大、成本高使应用受到限制;微透镜阵列方法是以牺牲图像的空间分辨率换取角度分辨率。因此采用掩膜方法进行高分辨率光场采集与重建,并以重建光场图像为基础完成深度估计。本文的主要研究内容如下:(1)研究压缩感知的基本原理与掩膜光场相机的对应关系,把压缩感知原理应用到光场重建中。压缩感知理论中信号的稀疏表示是能够重建的前提,为此详细阐述了K-SVD算法原理,用K-SVD算法训练光场样本集获取光场过完备字典,满足光场图像的理想稀疏表示,以更好的重建光场图像。(2)对随机测量矩阵优化,满足光场的物理重建需要。对优化的随机测量矩阵进行仿真光场图像的采集与重建,证明此方法可以重建高空间分辨率和大角度分辨率的光场图像。在此基础上搭建基于掩摸的物理光场采集平台,研究真实采集平台下物理掩摸-投影矩阵-测量矩阵的转换关系,完成掩摸到测量矩阵变换。最后通过光场过完备字典、掩膜转换得到的测量矩阵、CCD编码采样图结合压缩感知重建算法,实现真实光场的物理采集和重建。为真实光场图像的获取提供一种简单、有效的方法。(3)分析光场图像重聚焦原理,用光场角度像素块移动求和替代复杂积分实现光场图像重聚焦。为更好的估计遮挡边缘的深度信息,本文先通过分析研究光场图像特点,对存在遮挡部分的角度像素块分割优化,以解决深度估计时光场图像遮挡的问题。然后采用边缘轮廓信息优化深度图,相比遮挡线索优化深度图的方法,本文方在保证计算精度同时降低了算法的复杂度。最后将重建的光场图像进行重聚焦与深度估计。
[Abstract]:The light field image contains abundant spatial 3D information, so the light field image can be used for refocusing, depth estimation and 3D display, in which accurate depth information is used to detect salience, super-resolution reconstruction, etc. The development of target recognition and 3D surface reconstruction plays an important role. Although the method of obtaining light field image of camera array and microlens array can record the light field image effectively, the application of camera array method is limited because of its large volume and high cost. The method of microlens array is to sacrifice the spatial resolution of the image for the angular resolution. Therefore, the high resolution light field acquisition and reconstruction are carried out by mask method, and the depth estimation is completed based on the reconstructed light field image. The main contents of this paper are as follows: (1) the relationship between the basic principle of compression sensing and the mask light field camera is studied, and the principle of compression sensing is applied to the reconstruction of light field. The sparse representation of signal in compressed sensing theory is the premise of reconstruction. The principle of K-SVD algorithm is expounded in detail. The K-SVD algorithm is used to train the sample set of light field to obtain the over-complete dictionary of light field to satisfy the ideal sparse representation of light field image. The random measurement matrix is optimized with better reconstruction of light field image to meet the physical reconstruction needs of light field. The acquisition and reconstruction of simulated light field images based on the optimized random measurement matrix show that this method can reconstruct high spatial resolution and large angle resolution light field images. On this basis, the physical light field acquisition platform based on mask is built, and the conversion relationship between physical mask and projection matrix and measurement matrix is studied under the real acquisition platform, and the mask to measurement matrix transformation is completed. Finally, the physical acquisition and reconstruction of the real light field are realized through the over-complete dictionary of light field and the measurement matrix CCD coded sampling image obtained by mask conversion combined with the compressed perceptual reconstruction algorithm. This paper provides a simple and effective method for obtaining real light field image. It analyzes the principle of light field image refocusing and realizes the refocusing of light field image by moving the pixel block of the light field angle instead of the complex integral. In order to better estimate the depth information of occlusion edge, this paper analyzes the characteristics of light field image, and optimizes the segmentation of angle pixel block with occlusion part in order to solve the problem of depth estimation of time field image occlusion. Then, the edge contour information is used to optimize the depth map. Compared with the method of shading cues to optimize the depth map, the computational accuracy is guaranteed and the complexity of the algorithm is reduced. Finally, the reconstructed light field image is refocused and depth estimation is carried out.
【学位授予单位】：郑州大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【相似文献】