早期視覺下之生物啟發式混合質感邊界偵測模型

Abstract

本論文提出一個由生物觀點啟發的多通道質感邊緣偵測演算法用以偵測不同質感間的邊界。此演算法用高斯濾波器抽取質感的一階特徵，另外使用一組不同參數的GABOR濾波器抽取二階特徵。這些不同的特徵接著被予以整合形成一個N維的特徵空間。我們會分別計算每一個像素點(pixel)和他相鄰點在特徵空間中的差異，在消除調差異小的像素點之後，我們可以得到一個粗的邊界影像。最後我們使用區域頂點偵測的方式，找出精確邊界的位置。 此演算法簡單並且直觀，因此可實現於仿細胞神經網路(Cellular Neural Networks; CNN)。CNN擁有一些重要的特性，例如有效率的及時運算能力及方便於大型積體電路(VLSI)的實現。 在論文裡，我們大量測試我們的演算法在合成的質感影像上，而這些質感都是隨機從“Brodatz texture”中取出來的。由實驗結果我們可以發現均勻質感的邊界都可成功而精確的找出，而對於不規則或不均勻的質感，演算法仍會找出一些符合我們人眼感受的特性。

In this thesis a multi-channel texture boundary detection technique inspired from human vision system is presented. This algorithm extracts 1st-order features by a Gaussian filter and 2nd –order features by a set of even-symmetric Gabor filters. The hybrid-order features are integrated to construct an N-dimensional feature space. The difference between each pixel with its neighbors is measured in feature space, and coarse boundaries are obtained after eliminate pixels with small difference. After obtaining coarse boundaries, we use local peak detection to get the precise boundaries. The proposed algorithm is simple and straight-forward such that it can be implemented on Cellular Neural Networks (CNN) which possesses some important characteristics such as efficient real-time processing capability and feasible very large-scale integration (VLSI) implementation. We also extensively tested our algorithm on synthetic textures randomly picked from “Brodatz texture”, and from experiments it can be found that boundaries of uniform textures are detected successfully and have high spatial-accuracy. For the textures that are non-uniform or non-regular, the results also reflect some meaningful properties that consistent to human visual sensation.