使用遺傳演算法來設計模糊類神經網路中的特徵擷取模組

Abstract

以往在以類神經網路為基礎的辨識系統中，辨識率是由人類專家擷取的特 徵來決定的。此種系統的架構因需要多層的轉換而相當複雜。圖型辨認在 許多的應用中已是不可缺少的一部份，因此，全自動化辨認是目前研究的 重點。一種稱為 模糊濾波類神經網路 ，可不斷的學習並自動偵測特徵 的架構，已被成功地運用在電漿光譜分析上。在此論文中，我們將之擴展 至解決手寫體數字辨認，以顯示其一般性。一維模糊濾波器、二維模糊濾 波器、和以遺傳演算法為基礎的模糊濾波器，皆可用來做特徵擷取，且此 三種版本皆可很順利地以自動化的方式來解決在真實情況下的辨認問題： 偏移誤差與雜訊。實驗結果顯示這種架構有很高的辨識率，並具有在學習 效率上的優點。 In NN-based recognition systems, the recognition rates dependent on the quality of feature extraction which is usually determined by human experts and the models are very complex because they need multi-layered transformation. Since pattern recognition is an essential part in many applications, automating this task becomes more and more important. A neuro- fuzzy model of adaptive learning and feature detection, called the fuzzy-filtered neural networks, has been successfully applied to the problem of plasma spectrum analysis. In this thesis, we extend the model to another problem, the recognition of hand-written numerals, to demonstrate its generality. We proposer three versions of the architecture, which use one- dimensional fuzzy filters, two-dimensional fuzzy filters, and genetic-algorithm-based fuzzy filters, respectively, as feature detectors. All three versions smoothly and automatically handle issues of a real-world pattern recognition problem such as drifting and noise. Simulation results show that the proposed model is an efficient architecture for achieving high recognition accuracy.