光場複眼鏡頭應用於機械視覺之場景深度量測方法

Abstract

我們利用積分攝影(integral photography)的概念取得四維光場(light field)，並以此為基礎提出還原景深的演算法。四維光場不同於以往傳統相機所記錄到的影像資訊，他可以記錄到光進入相機時每束光線的方向，並且被使用在場景深度估測、重新對焦等應用上。本研究由市售的光場攝影機Lytro做為影像的來源，其內置入了微透鏡陣列(micro-lens array)只要拍攝單張照片便可得到各個視角的資訊，因此可以利用影像幾何的三角關係將原圖重新排列合成Epipolar plane image，再加以分析得到景深估測的結果。而微透鏡陣列的大小排列情形會影響到四維光場區分為方向解析度和空間解析度時的組成，其中因為Lytro原始的光感測器尺寸和解析度規格較低，使最後擷取出來的方向解析度為7*7，空間解析度為300*300，在原始資訊較不充足的情況下，我們在提出的演算法中加強相關高低階影像處理例如:物體分離(segmentation)可以在進行景深估測前預先知道每個物體的存在，使在估測當下減少誤差的產生;非線性結構張量(nonlinear diffusion)改善了傳統的高斯濾波器造成邊緣模糊的情況，加強物體邊緣資訊的保存，以上方法都使場景深度的估測結果達到較好的抗雜訊效果。

4D light field offers the necessary information for scene reconstruction. It is thus desirable to use it as a quick mean to obtain three dimensional scene. In this thesis, we first use the concept of integral photography to get the 4D data from the scene. We then propose an algorithm for the depth recovery. The 4D light field is taken with a commercial plenoptic camera from Lytro. The plenoptic camera uses an added micro-lens array which allows us to also record the direction of the incoming beam. Although Lytro uses the lens array information for refocusing, here we use it for depth recovery. The depth information is recovered by transforming the raw image into an EpiPolar-Image (EPI) and then apply triangular geometry to obtain a preliminary result. This preliminary result is basically too noisy for object extraction. We then apply object segmentation and nonlinear diffusion to achieve useful crisp 3D scene information. Due to the trade-off between directional resolution and spatial resolution, the spatial resolution is only 300*300.