神經網路於影像切割及控制系統之應用

Abstract

本文分為兩部份。第一部分我們探討神經網路於影像切割上的應用。我們 採用生物視覺模型導出之神經網路－FBF網路以利用人類生物視覺系統之 優越特性來解決傳統影像切割之方法無法處理好的問題。首先，我們將探 討FBF網路中最重要之元素－並列競爭式神經網路(shunting competitive neural networks)之特性。其次，我們將說明一個生物視覺 模型－FACADE模型。再來，我們將介紹FBF網路之特性多方向性遮罩( multimask formulation)及階層式架構(Hierarchical structure)，並以 向量分析說明有方向性遮罩(Oriented masks)之特性。最後，我們提供了 許多例子說明網路參數之影響。第二部份的主旨在於探討高斯型類神經網 路在控制系統上的應用。我們將分析高斯型類神經網路之學習特性及其應 用於非線性控制系統之表現。另外，由於高斯型神經網路控制器與CMAC在 許多方面類似，我們將對此兩者做一比較。最後，我們將說明高斯型神經 網路控制器的學習時間及記憶體的需求將隨系統輸入及輸出維度而急劇增 加。 This thesis is divided into two parts. Part I of this thesis studies the biologically derived image processing techniques－ the FBF network. We take advantage of biological vision system to solve a lot of tradeoff existing in traditional image segmentation algorithms. We discuss the most important element of FBF network－the shunting competitive neural networks. A biological perception model which is called FACADE theory is discussed. The main point of part I is to discuss the structure and properties of the FBF network.The FBF network uses multimask formulation and hierarchical structure to detach a desired figure from background. A vector analysis for the oriented masks of FBF networks is provided. Finally,we provide some examples to explain the effect of parameter variation. Part II of this thesis concentrates on the application of Gaussian neural networks on nonlinear control system. A Gaussian networks design procedure, which determines the networks parameters via frequency analysis is reviewed. We discuss the performance of the Gaussian network controller and compare it with a similar neural network controller－the CMAC. Finally, we demonstrate that the training time and memory needed for achieving certain accuracy would increase tremendously as input and output dimension increase.