臺灣博碩士論文加值系統

手寫數字辨識具有很高的研究價值，其應用範圍非常的廣泛，可應用到
成績自動登錄系統，車牌資料檢查，郵遞區號辨認等等。但手寫字常因個人書
寫習慣不同，造成字形變異甚大。因此，多年來已有許多學者相繼提出辨識的
方法和設計辨識系統。
本論文將提出以監督式多維矩形複合式類神經網路（Hyper Rectangular
Composite Neural Network，簡稱HRCNN）為基礎之手寫數字辨識並應用於成績
自動登錄系統。成績自動登錄系統分為三部份：前處理部份、定位部份及辨識
部份，分別處理考卷上的手寫數字及印刷體數字。第一部份前處理包括影像的
取得、圖文分離、影像二值化及分數數字的切割。第二部份是利用連接區塊的
分析，去除分數切割後多餘的筆劃及找出准考證號碼在影像中的位置。第三部
份則是數字辨識部份，使用非線性正規化(Nonlinear Normalization)其目的是
使手寫數字正規化到一定的大小，且在筆劃的密度上做一適當的調整，及使用
局部化弧形樣本模型(localized arc pattern)的方法去抽取特徵(feature
extraction)參數，再輸入到多維矩形複合式類神經網路，最後辨認出結果。
本系統收集了70個人的手寫數字，每個人寫0-9的阿拉伯數字6組，其中3
組當訓練網路用，另3組當作closed test，另外再收集了20個人的手寫字當open
test用（不包括前面的70個人）獲得不錯的效果。最後再測試了4位老師各20份的
考卷共80份考卷作open test，在准考證號碼的辨識由於是印刷體字所以高達
100%的辨識率，考卷分數的手寫數字也能達到93.75%的辨識率。

Handwritten numeral recognition has high potential in some applications in our daily life. It can be used in a wide range of applications, such as an automatic score register system, license-plate data verification, ZIP code recognition, etc. As a result, in the recent years many researchers have proposed relevant methods and systems for handwritten numeral recognition.
In this paper, the author proposed a handwritten digit recognition system based on a supervised HyperRectangular Composite Neural Network (HRCNN) and then applied this system to an automatic score register system. This system is composed of three parts: preprocessing, numeral extraction, and recognition and the author used this system in both the handwritten scores and the printed serial numbers on the examination paper. In the first stage, the image of the paper is obtained as the input and some processing is performed on the input image, such as image binarization and segmentation. In the second stage, object labeling is used to extract the connected components in the image. The connected components can be used to find the position of the serial number. In the third stage, nonlinear normalization is performed to get a normalized image for recognition. The purpose for using nonlinear normalization is to get a image with a fixed size and to adjust the density of the strokes in a adequate manner. The features using localized arc patterns are extracted from the normalized image. The features are then used as the input to the HRCNN and the recognition result can be obtained.
Handwritten numerals of 70 persons were collected as the data set. Each person wrote numerals from 0 to 9 six times. Three times of these are used as the training set and the others as the testing set. A good result was obtained for this data set. Another 80 examination papers were used for testing. These papers were collected from four teachers and each teacher provided 20 papers. The recognition rate in the serial numbers is 100% since the numerals are printed numbers. On the other hand, in the handwritten scores, a recognition rate of 93.75% was obtained.