基于低秩矩阵分解的光场稀疏采样及重构

光场成像技术中光场的采集和数据的压缩处理是亟待解决的问题。为了实现光场的稀疏采样和恢复,建立了基于光场低秩结构的压缩采样相机系统,研究了光场矩阵的结构特征及压缩采样下光场图像的重构问题。根据静态光场各视点图像之间的内容相似性,将这些图像向量化并按列组合成一个二维矩阵,该矩阵呈现出低秩或近似低秩的状态。对光场图像矩阵进行低秩分解,结果表明偏离低秩的部分呈现出很强的稀疏性性质,低秩和稀疏各自表征不同的数据冗余度。然后,对基于掩膜的相机采样系统进行随机Noiselets变换测量,鉴于重构过程是一个低秩稀疏相关性约束下的优化求解问题,采用贪婪迭代求解分别重构出光场矩阵的低秩部分和稀疏部分。仿真结果表明,重构图像的PSNR维持在25dB以上,且保留了光场视点间的视差信息,能够满足稀疏采样中对光场图像的要求。

Sparse sampling and reconstruction of compressive light field via low-rank matrix decomposition

Collection of light field and compression of data in light field imaging technology are urgent problems which need to be solved.In order to realize sparse sampling and restoration of the light field, a camera system to compress samplings based on low-rank structure of the light field was built for researching structural features of matrix of the light field and the reconstruction of light field images under compressive sampling.According to content similarities between each viewpoint image in static light field, those images were vectorized into a two-dimensional matrix by columns.The matrix presented a low-rank or approximated low-rank state.Low-rank decomposition of image matrix in the light field were finished, which shows that deflective low-rank parts emerge strong sparse properties, and low-rank and sparseness separately represented different data redundancies.Then, the camera sampling system fitted with the mask was measured through sparse random Noiselets conversion.Considering the reconstruction process was an optimization solution problem constrained by low-rank sparse correlation, the greedy iterative solution was adopted to separately reconstruct low-rank parts and sparse parts of light field matrix.The simulation result shows that the PSNR of reconstructed image that keeps disparity information among viewpoints of the light field maintains over 25 dB, thus meeting the requirement of sparse sampling for images of light field.