智慧型立體對應式之視差法則設計

Abstract

在本篇論文中，提出處理一對左右影像特徵點立體對應問題的演算法來偵測視差，而特徵點是使用Harris角落偵測器所抓取而得到的。此立體對應問題可被系統化為一個最佳化問題的目標函數，該函數代表問題答案的限制和特性，而且此目標函數可轉換成二維Hopfield神經網路的能量函數去做最小化處理。Hopfield神經網路是一個單層回授網路，每個神經元代表左圖與右圖各一點的對應關係，當網路裡所有的神經元輸出皆不再改變時，也就是網路達到穩定狀態，此時可獲得對應結果和對應配對之視差。本研究比較了三種不同之目標函數的效能，並用一個簡單的應用來呈現視差的偵測。

In this thesis, an algorithm is proposed to detect the disparity by solving the stereo matching problem for a set of feature points extracted by the Harris corner detector from a pair of stereo images. The stereo matching problem is formulated as an objective function which represents the constraints and properties on the solution. Then the objective function is transferred to the energy function of 2D discrete Hopfield neural network for minimization. This neural network is a single layer feedback network and each neuron in the network represents a possible match of two feature points, one in the left image and the other in the right image. The matching result and the disparity of the matched pairs are obtained when every output of neurons are no longer changed, i.e. the network reaches its stable status. Furthermore, the performances of three kinds of objective function are compared, and a simple application is presented to show the disparity detection.