以自適應類神經網路為基礎之字符辨識

文化大學機構典藏 CCUR > 理工學院 > 工學院 > 機械工程系暨機械工程學系數位機電研究所 > 博碩士論文 > Item 987654321/25511

jsp.display-item.identifier=請使用永久網址來引用或連結此文件: https://irlib.pccu.edu.tw/handle/987654321/25511

题名:	以自適應類神經網路為基礎之字符辨識 Character Identification Based on Adaptive Neural Network
作者:	廖浩宇 Liao, Hao Yu
贡献者:	機械工程學系數位機電碩士班
关键词:	自適應類神經網路梯度倒傳遞類神經網路網路仿真字符辨識 Adaptive Neural Network Gradient Backpropagation Neural Network Network Simulation Character Recognition
日期:	2012-12
上传时间:	2013-10-14 11:38:04 (UTC+8)
摘要:	本研究提出自適應類神經網路演算法來做字符辨識，網路訓練是以非線性的映射方式，依據分類的誤差值以及調變網路中的學習率α值達到快速收斂，辨識能力提升的目標。樣本共有36個32X32大小的圖樣，分別是0~9的數字圖樣及A~Z的英文字圖樣，首先，找出36個圖樣的像素值設定為P值當作輸入層，再把36圖樣裡面的數字和英文字設定為T值當作輸出層，本研究使用 MATLAB 的類神經網路工具箱做為開發工具，從設定訓練函數、適應性學習函數、性能函數、輸入層、輸出層、動量項係數、網路層數、學習率α、學習率遞增、學習率遞減、性能遞增最大值、神經元數目，去訓練以倒傳遞類神經網路為主，輔以自適應學習網路的字符辨識架構。從實驗結果，顯示可到達相當好的辨識能力。研究中也與梯度倒傳遞類神經網路的字符辨識方法做比較，結果顯示本研究所提出的方法，具有較快的收斂性及較佳的辨識能力。此項研究架構共分三步驟，分述如下：步驟一：將0~9和A~Z的圖樣建立到資料夾，以MATLAB讀取字符圖樣，數字、英文字，進行網路訓練。步驟二：利用MATLAB讀取字符圖樣，數字、英文字，進行網路仿真的訓練，訓練出的出P值和T值，當作Input和Output。步驟三：利用MATLAB撰寫類神經網路機制，並且利用MATLAB裡面的工具箱去設定網路的訓練函數、適應性學習函數、性能函數、輸入層、輸出層、網路層數、學習率α、學習率遞增、學習率遞減、性能遞增最大值、神經元數目，以非線性的方式，將36個圖樣訓練出來，進行梯度倒傳遞和加入自適應類神經，比較那種類神經的收斂性能力較快，最後，加入高斯雜訊和中位數濾波辨識圖樣，看哪種辨識能力較佳。本研究已成功的利用MATLAB、梯度倒傳遞類神經網路及自適應類神經網路，完成了多種圖樣之辨識。本研究之基本成果，日後可再增強影像前置處理機制，以增強後續的辨識能力，也可大幅提升實務應用，以解決目前服務型字符辨識系統上之盲點，亦可提升執行上的效率，圖樣的準確性。 This thesis proposes an effective adaptive neural network algorithm to recognize the character pattern. In order to achieve rapid convergence and increase the identification capability, the learning rate α of the neural network is modulated according to the classification error of previous iterations. The total 36 samples include 10 digital patterns (0-9) and 26 character patterns (A-Z). Every sample contains 32X32 size of the pattern. Firstly, the pixel values of 36 patterns are found and set as the P-value of the input layer, then the represented digital values of 36 patterns are set as the T-value of the output layer. The neural network toolbox of MATLAB is utilized as the development tool in this study. From the settings of training function, adaptive learning function, performance function, input layer and output layer, momentum coefficient, number of layers, learning rate α, learning rate increasing factor, learning rate decreasing factor, maximum performance increasing value and number of neurons, an adaptive learning network architecture is developed to train a back-propagation neural network to recognize the character patterns. From the experimental results, the recognition ability and recognition speed of the proposed architecture are superior to the pure momentum neural network. The researching steps are divided into three parts, as shown below: Step 1: 0-9 and A-Z patterns are established and folded into MATLAB work files. Step 2: Start the MATLAB to read the established character patterns and to train the P-values and T-values of the input and output layer of the network. Step 3: Programming the MATLAB neural network mechanism by utilizing the toolbox to set the training function, adaptive learning function, performance function, input layer and output layer, momentum coefficient, number of layers, learning rate α, learning rate increasing factor, learning rate decreasing factor, maximum performance increasing value and number of neurons. All the 36 pattern are trained by pure back-propagation momentum neural network and the proposed adaptive neural network, respectively. Then compare the convergence curve and recognition speed. Furthermore, adding Gaussian noise and median filter to the patterns to compare the performance of above two architectures. This study has successfully used MATLAB, back-propagation momentum neural network and adaptive neural network to implement the patterns recognition. In the future, the character recognition could be further enhanced by embedding more pre-processing mechanism. And the performance of blind spot can be improved in practical application.
显示于类别:	[機械工程系暨機械工程學系數位機電研究所] 博碩士論文

文件中的档案:

没有与此文件相关的档案.

检视Licence

在CCUR中所有的数据项都受到原著作权保护.

数据加载中.....