Surface Reconstruction Based on Plenoptic Image with Convolution-Based Iterative Algorithm

Abstract

Plenoptic imaging is a revolutionary photographic technique which has been brought to public awareness recently. It makes post-refocusing possible by capturing both spatial and directional information of light. In order to do that, a micro-lens is placed before the imaging plane of the camera. Each micro-lens records angular information from its location at the cost of spatial resolution. A focal stack is generated by stacking up refocused images of the same object. By adjusting the depth at which each image focuses, it models the light ray around the nominal focal plane of the camera in 3D. The main purpose of this thesis is to reconstruct the surface based on focal stacks. There were previous attempts to accomplish the same task, including the gradient method and the stereo method; however, they are unable to reconstruct a smooth object. Later research proposed a deconvolution method that can address the smooth object problem. The thesis demonstrates the limitation of the deconvolution method and proposes an iterative method to more precisely reconstruct the surface of the object. In order to implement the iterative algorithm, the thesis also mathematically models the process of image blur as the vectorized object multiplied with a convolution matrix which represents the point-spread function (PSF). The PSF describes the spread of light from the object across the entire focal stack. Various results of estimating depth from different methods are presented to compare the performance of each algorithm.