應用神經網路於 PSRI 圖樣識別之研究

Abstract

我們提出了一個架構能夠正確的辨識某一物體，而不會受到該物體在二維 空間上的位置改變､大小改變､或是平面上旋轉等的影響。在本論文中， 我們的目標是維持在高的正確辨識率下，能夠增加辨識的類別和提高影像 的解析度。我們提出的架構包含四個部份：粗糙編碼､特徵抽取､特徵正 規化和分類。粗糙編碼是為了提高辨識影像的解析度。特徵抽取是以兩個 三角形之間的相似度為量測標準。而兩個三角形之間相似的程度是由球形 參數來決定。我們提出兩種不同的特徵正規化的方法以配合我們所使用的 分類器－ Fuzzy ARTMAP。在模擬中，我們使用兩百筆不同圖樣的特徵來 訓練它，使用三千筆不同圖樣的特徵來測試它。這架構能夠辨識四類不同 的圖樣而不會受到二維空間上的位置改變､大小改變､或是平面上的旋轉 的影響。由模擬的結果發現：在相同的四類圖樣測試下，其正確辨識率比 高階神經網路和回傳神經網路好很多，但其雜訊容忍度較高階神經網路差 。 The proposed algorithm could fully recognize an object despite changing in the object's position in a two-dimensional field, changing in its size, or in-plane rotation of the object. In this thesis, we aimed to increase the number of classes recognized and the image resolution of the input field while maintaining a high level of accuracy in recognition. The algorithm included four components: coarse coding, feature extraction, feature normalization, and classification. Coarse coding was used for increasing the resolution of the recognized input field. The feature extraction was used to measure the similarity between two triangles. The degree of similarity between two triangles is measured using the parameter- sphericity. We proposed two different methods of feature normalization in this thesis. Fuzzy ARTMAP was used as the classifier of the algorithm. Our simulation show that the recognized accuracy of the algorithm was superior to that of higher-order neural network and backpropagation neural network, but noise tolerance of the algorithm was more poor than that of the higher-order neural network.