類神經網路使用於強健的辨認

Abstract

在本論文中,我們提出類神經網路使用於強健的辨認.霍氏函數使用來做特 徵的抽取.幅射函數網路被訓練當作一個分類器.它的學習法則分為兩個步 驟.在隱藏層學習法則,一個分群法則用來決定隱藏層的節點個數和每個節 點的中心.在輸出層的學習法則,使用最小誤差的學習法則.在強健辨認對 抗雜訊方面,我們考慮幾個步驟,每個步驟在網路學習之後使用來辨認有不 同程度雜訊的圖形, 完全沒有雜訊的圖形上且隨步驟進行,逐漸加大此雜 訊. 形加入原來的圖形中,網路再重新學習.重新學習的網路辨識能力會被 改善.在實驗上,此方法應用於印刷體中文字及震測圖形辨識問題上,而這 些結果都相當不錯. In this thesis, we propose the robust recogniton using neural network. The Walsh function are emplued for feature extraction. Radial basis function network is trained as the classifier. Its learning algorithm is divided into two stages. In learning hidden layer, a clustering algorithm, is determined the number of hidden nodes also them of "centroid". In learning output layer, a least mean square algorithm, superbised learning, is used. In the robust recognition against noise, we consider several steps. Each step, after the network trained with original training patterns, the network is used to classify a set of noise patterns which generated by adding some degree of noise into the original training patterns. Then the misclassified patterns are added into the original training patterns set for relearning. After relearning the recognition rate can be improved. In the experiments, the printed Chinese characters and the seismic pattern recognition problem are applied. And the robust recognition results of those experiments are good.