利用類神經網路進行平行細線化及結構性特徵抽取並進行手寫中文字辨認

Abstract

本論文提出一使用類神經網路辨識手寫中文字的方法。在本方法中，包括 影像細線化、結構性特徵抽取、及文字分類等辨識中文字所需的所有處理 步驟均由類神經網路處理。整個系統包含幾個互相連接的類神經網路。首 先將一手寫中文字影像輸入一細線化類神經網路進行細線化。接著，將細 線化結果送入一特徵抽取類神經網路，進行結構性特徵抽取。最後，由一 文字辨識類神經網路利用這些結構性特徵及其位相關係，對所輸i入之手 寫中文字作辨識。除了這些類神經網路外，本論文並提出一單迴平行細線 化演算法 OPPTA。本論文共提出三個細線化類神經網路，這三個類神經網 路均植基於單迴平行細線化演算法 OPPTA 。此細線化演算法利用樣型比 對將物體的邊緣點一層一層地剝離，來產生一完美八連接、不受雜訊影響 、且無過度剝離現象之細線化結果。因此一演算法在單一輪迴 (pass) 中 就將所有的邊緣點一次剝離，故可直接依其設計產生細線化類神經網路。 以細線化演算法 OPPTA 為基礎，本論文所提出之三個細線化類神經網路 的第一個是含有三層架構的迴歸類神經網路 (recurrent neural network)，此類神經網路由非常簡單的處理元 (processing element)組 成，因此結構非常龐大。在修正處理元的輸出函數 (output function)後 ，本研究設計出了第二個細線化類神經網路，此類神經網路為一含兩層架 構的迴歸類神經網路。第三個細線化類神經網路是一單層架構的類神經網 路。此項簡化導因於允許各處理元以 sigma-pi 函數來收集其輸入訊號接 著，本論文提出一利用細線化類神經網路結果進行特徵抽取的結構化特徵 抽取類神經網路系統。此網路系統由兩個相接續的類神經網路組成，即直 線分割網路 (line separation network, LSN) 與特徵抽取網路 (feature extraction network, FEN)。 LSN 依線的走向將線上的點分成 四類，分類結果分別表現於代表四個走向的四個神經元平面 (neuron plane)。 FEN 則從這四個神經元平面中抽取結構性特徵。 An integrated scheme for recognizing handprinted Chinese characters is proposed, in which all the processing stages required to recognize a Chinese character are performed by cascaded neural networks, including neural networks for image thinning, structural feature extraction, and character classification. At the beginning, the image of a handprinted Chinese character is fed into a thinning neural network. The thinning result is then sent to a feature extraction neural network system to derive the structural features. Finally, a character recognition neural network recognizes the handprinted Chinese character by the extracted structural features and the topological relationships among them. Three neural networks are proposed for image thinning. All the neural networks are based on a new one-pass parallel thinning algorithm called OPPTA, which is also proposed in this dissertation study. Algorithm OPPTA removes boundary points layer by layer by matching a set of templates with an input binary image and produces perfectly 8-connected and noise-insensitive results without excessive erosion. Since this algorithm removes all boundary pixels in a single pass, neural networks for image thinning can be implemented directly from it. The first of the three neural networks proposed for thinning binary images is a three-layer recurrent neural network. Being constructed by simple processing elements, this neural network is quite huge in size. By changing the output functions, a two- layer simplified version of the first neural network for image thinning is obtained. The third neural netwok for image thinning is a single layer neural network. This simplification is achieved by introducing the capability of performing the sigma-pi function of collecting inputs into the processing elements.