以類神經網路為基礎之適應性整合式反射模型及其在三維立體重建之應用

Abstract

本論文提出一個以類神經網路為基礎的適應性整合式反射模型及其在三維立體重建之應用。這個類神經網路自動整合光學成像上的散射與反射成分，使得我們可以個別考慮物體表面上每一點的成像特性，並且針對表面不同反射率的問題加以處理。我們將二維影像輸入到多層級的類神經網路，透過指導性學習法，可以得到物體表面每一點的法向量。接著利用”Enforcing Integrability”方法，我們從計算得到的法向量重建出物體的形狀。為了測試我們的方法應用在人臉與其他物體的效果，我們設計一個多光源的拍攝環境。為了使被拍攝物體從各種角度所得到的光源強度一致，我們將拍攝環境設計成一個半圓球。為了使取像的動作和閃光燈亮的動作同步，我們設計了一個控制板來控制閃光燈。有了這樣的拍攝環境，我們可以輕易地改變光源的角度並且取得在這些光源下拍攝的影像，方便實驗的進行。我們做了四種實驗來比較本論文的方法與過去方法的重建效果。另外，我們探討光源位置對重建的影響以及輸入二維影像的多寡對重建效果的影響。本方法具有以下優點：(1) 經由類神經網路的學習能力，改進過去必須事先知道光源位置的缺點。(2)考慮物體表面上各點不同的反射特性，而不是將每一點的特性都視為一樣。(3)處理表面不同反射率的問題，避免重建時造成失真。(4)經由實驗得知，我們的方法可以應用在較多的物件上而且可以得到較佳的重建結果。

In this thesis, a neural-network-based adaptive hybrid-reflectance model is proposed for 3-D surface reconstruction. The neural network combines the diffuse component and specular component into the hybrid model automatically. We can consider the characteristic of each point individually and solve the problem of variant albedo. The pixels of the 2-D image are fed into the multi-layer neural network and we can obtain the normal vectors of the surface through supervised learning. Then enforcing integrability method is used for the reconstruction of 3-D objects from the obtained normal vectors. In order to test the performance of our proposed algorithm on the facial images and other images of general objects, we design and construct a photographing environment to satisfy the requirement of the proposed method. To make the strength of different light sources illuminating to the photographed objects equal, the photographing environment is constructed as a hemisphere structure. In order to synchronize the capturing action with the trigger of the electronic flashes, we also design a control board to control the electronic flashes. With this photographing environment, we can modify the direction of illuminant source easily and capture images under variable illumination in a short time. Finally, four experiments are performed to demonstrate the performance of the proposed method. The advantages of our method are summarized as follows: (1) By the learning ability of neural network, we don’t need to know the illuminant direction in advance. (2) The individual characteristic of each point on the surface is considered. (3) The problem of variant albedo is considered to avoid the distortion of surface reconstruction. (4) According to the experimental results, our neural-network-based adaptive hybrid-reflectance model can be applied to more general objects and achieve better performance for surface reconstruction.