【正文】 CT Today the science and technology develop rapidly. Letter recognition technology based on the feedback neural work is applied in many aspects including publication, finance military, cash register, page views, and any with repeatability,and variability of data files . Letter Identification System include the following processes: preprocessing, feature extraction, BP neural work training,and recognition.. In this paper, we use a threelayer neural work, including input layer, hidden layer and output layer. This paper supply of a variety of methods to determine Hidden layer nodes . The root sign method and other proposed by the Nelson and Illingwnrth are applied . The features and advantages of Artificial neural work is reflected in three aspects : First, a selflearning function. When we recognize letters, only putting many different images and the corresponding results into the artificial neural work and forming a stable weight before the letter recognition,the work will be through selflearning function to slowly identify similar , with the association storage. Artificial neural work feedback work can achieve this association in the letter recognition. Third, finding the optimal solution with high capacity. Finding the optimal solution of a plex often require a large amount of a design that a feedback type artificial neural work for problem and playing the highspeed puting power of puter, you may quickly find the optimal solution. In the matlab environment this article simulate the entire process of letter recognition, with the development of science and technology recognition technology is more mature and have various problems will be solved. Keywords： Letter identification； image processing； feature extraction； the feedback neural work 濟南大學泉城學院畢業(yè)論文 III 目 錄 摘 要 ........................................................................................................... 1 ABSTRACT .......................................................................................................... II 1 前言 .................................................................................................................. 1 研究背景及意義 ..................................................................................... 1 研究現(xiàn)狀 ................................................................................................. 2 手寫字母識別方法 ................................................................................. 3 結構模式識別方法 .................................................................... 3 統(tǒng)計模式識別方法 .................................................................... 3 統(tǒng)計與結構相結合的識別方法 ................................................ 4 人工神經(jīng)網(wǎng)絡方法 .................................................................... 4 識別系統(tǒng)性能的評價 ............................................................................. 5 論文組織結構 ......................................................................................... 5 2 預處理 ................................................................................................................ 6 系統(tǒng)框架 ................................................................................................. 6 預處理概述 ............................................................................................. 6 本文預處理設計 ..................................................................................... 6 去噪 ............................................................................................... 7 二值化 ........................................................................................... 8 歸一化 ...................................................................................... 10 細化 ............................................................................................. 11 3 字母特征提取 ................................................................................................ 13 特征提取概述 ....................................................................................... 13 本文特征提取設計 ............................................................................... 13 像素百分比特征 ......................................................................... 14 提取矩陣的粗網(wǎng)格特征 ............................................................. 15 重心特征 .................................................................................... 16 提取圖像的矩陣像素特征 ........................................................ 16 筆劃特征 ..................................................................................... 17 外輪廓特征提取 ........................................................................ 18 4 BP 神經(jīng)網(wǎng)絡 .................................................................................................. 19 人工神經(jīng)網(wǎng)絡 ....................................................................................... 19 神經(jīng)網(wǎng)絡的模型圖 ................................................................................ 20 BP 神經(jīng)網(wǎng)絡的工作原理 ................................................................... 21 濟南大學泉城學院畢業(yè)論文 IV 神經(jīng)網(wǎng)絡的各層節(jié)點數(shù) ........................................................................ 22 輸入層和輸出層 ....................................................................... 22 隱含層節(jié)點數(shù)的優(yōu)化確定 ...................................................... 23 BP 神經(jīng)網(wǎng)絡的參數(shù)設計和訓練過程 [17] ............................................. 25 5 實驗結果及分析 .............................................................................................. 28 實驗設計 ............................................................................................... 28 實驗參數(shù) .................................................................................... 28 訓練和識別樣本庫設計 ............................................................ 28 隱含層節(jié)點對實驗結果的影響 ............................................................ 28 識別樣本的正確率 ................................