本文利用啟發式類神經網路做圖樣辨識，類神經網路是以非線性的映射方式，將輸入的特徵值對應到網路輸出分類的結果，並依據分類的誤差值以及調整網路中的加權值使其達到收斂狀態，讓辨識能力提升。本研究以倒傳遞類神經網路為主輔以啟發式學習方法來辨識圓形，X形，梯形圖樣。本研究用Matlab執行CCD鏡頭抓取圖形並用不變矩的方法算出圖形的特徵值，放入學習完成的類神經網路來進行辨識。
經過多次實驗可以看出啟發式學習方法在學習率的收斂和誤差平方值皆優於傳統倒傳遞的方法，結果顯示本研究所提出的方法，具有較快的收斂性和較佳的辨識率。此項研究架構共分四步驟，分述如下：步驟一：CCD鏡頭抓取我們要的圖樣圓形，X形，梯形；步驟二：利用不變矩算出每張圖樣的六個或七個特徵值；步驟三：用兩種不同類型的類神經網路分別訓練事先放入的36個特徵樣本；步驟四：捕獲並輸入待辨識圖型，放入分別已訓練完成的神經網路進行辨識。本研究已成功的利用Matlab、啟發式類神經網路，完成了多種圖樣之辨識。
The heuristic neural network (HNN) is utilized to recognize the pattern in this thesis. The nonlinear mapping between input and output is adopted in the neural network (NN). To achieve stable convergence and to promote the rate of identity, the neural network is trained by adjusting the weighting values between layers according to the gathered characteristic data and heuristic learning method. In this thesis, a heuristic learning based back-propagation neural network is proposed to recognize three patterns - ‘round’, ‘X’ and ’ladder’. Before recognition, the patterns are captured by CCD camera and then the characteristic eigenvalues are calculated in MATLAB environment. After gathering these characteristic data, the neural network is trained.
From the training experiment, the performance of HNN is better than that of the traditional back-propagation NN in the convergence of the learning rate and rate of identity. The simulation results demonstrate that the proposed method has faster convergence and better rate of identity. The developing stages are divided into four steps: Step 1: Capturing the patterns by CCD camera; Step 2: Calculating six or seven characteristic values via the variant moments; Step 3: Training the HNN by the gathered 36 eigenvalues; Step 4: Capturing and input new pattern. The pattern recognition has been implemented successfully by Matlab and the proposed HNN in this thesis.
