植基於WGLVQ離線式手寫數字辨識

Off-line Handwritten Numeral Recognition

李健宏

指導教授 ： 葉榮木 博士

國立臺灣師範大學/科技學院/工業教育學系在職進修碩士班/碩士(2002年)

摘要

離線式手寫數字辨識雖已被研究多年，但因為手寫字變異性大，對辨識研究者是極大挑戰；雖然手寫數字辨識已有高辨識率但仍有學習及辨識耗時間等缺點，本研究重點能改善上述缺點，但是仍要有高辨識率。
本研究採用MNIST資料庫做為訓練及測試用資料。特徵抽取，使用統計式特徵(Statistic feature)方法。雖然統計式特徵會有特徵向量維度很高的缺點，但本研究以有效特徵抽取方法，使特徵維度再130低維度就能將十種分類區分開。
為了提升辨識效果，在這研究中加入Fisher’s LDF (Linear discriminant function)特徵轉換，經實驗証實在不分群情況下，以特徵轉換後的特徵(未經學習訓練)做辨識就有92.6%的極高辨識率。本研究係以GLVQ(Generalized Learning Vector Quantization)為基礎，GLVQ為針對LVQ的收歛性與完整性加以改善。本研究除了探討LVQ及GLVQ理論並將其應用於手寫數字辨識，經實驗証明了兩者都有不錯辨識效果，也證實我們的特徵處理方法是有效的。
本研究中提出加了權重(weight)的LVQ及GLVQ成為WLVQ及WGLVQ，新學習法WLVQ及WGLVQ在每次學習過程中除了調整代表分類的參考向量外，同時也調整各分量權重，對權重輕的每次往下調的多點使權重輕的與權重重的權種值區分的更加明顯，經實驗証明WGLVQ或WLVQ都比原先分類器有較好辨識效果。
經過WGLVQ分類器學習訓練後以測試資料辨識(open test)會有97.6%辨識率，若再把每種分類分成16群，則更可將辨識率提高到98.2%，辨識率雖不及Ernst所提出的LIRA分類法有99.3%辨識率，但本研究辨識10000筆資料僅需1~2分鐘，相較於LIRA分類法虛耗時高達30分鐘，本系統應屬具有實用性。

關鍵字

手寫數字辨識 ； HASH(0xccedaec) ； HASH(0xccedb64) ； HASH(0xccedbdc) ； HASH(0xccedc54) ； HASH(0xccedccc)

並列摘要

Recognition of off-line handwritten numerals has been the subject of research for many years. Since handwritten numerals widely vary in their shapes, recognizing them has been difficult and challenging. Although a high level of recognition has been achieved, the shortcomings of time-consuming learning and recognition still persist. The present research focuses on overcoming these defects, while maintaining a high recognition level.
The research discussed in the present paper makes use of the MNIST database for learning and testing. For feature extraction, statistic features are used in the present research. Employing statistic features is saddled with the difficulty of a high number of dimensions, yet the present research, by using 130 dimensions, is able to distinguish between ten classifications.
To make character recognition more effective, in the present research transformation by Fisher's LDF (linear discriminant function) is applied to input characters. As experiments have shown, after transformation of non-clustered features (without learning) a level of recognition of 92.6% is achieved. In the present research, the method of WGLVQ, which is based on GLVQ (generalized learning vector quantization), is employed. Better convergence is achieved by GLVQ, and it is able to improve for LVQ. Experiments conducted within the current research have shown that both LVQ and GLVQ, applied to recognizing handwritten numerals, have quite good convergence behavior, also confirming the effectiveness of feature processing presented here.
In the present research, the methods of LVQ and GLVQ are enhanced by weighting, yielding novel methods of WLVQ and WGLVQ. Therein, in every learning step, not only directions classifying reference vectors are adjusted, but also weights of every vector. With every step the weights of less-weighted vectors decrease, resulting in more pronounced distinctions of light and heavy weights. According to experiments, both WLVQ and WGLVQ exhibit more effective character recognition.
With classification by WGLVQ and including learning, in an open test a level of recognition of 97.6% is achieved. With 16 clusters for each class, the recognition level rises to 98.2%. This result trails the level of 99.3% attained by Ernst using classification by LIRA, but while recognizing 10000 samples takes 30 minutes for the LIRA’s classification, the present approach allows recognition of 10000 samples in 1 - 2 minutes. The present research offers a more practical approach.

並列關鍵字

Numeral Recognition ； LVQ ； GLVQ ； WGLVQ ； feature transformation ； Fisher's

參考文獻

1.D.-F. Chang, J. Y.-J. Hsu, and C.-S Fuh. Handwritten character recognition using a neural network. 5th OCR & DA, pp. 17-20, Taiwan 1996.

2.G. M. T. Man, J. C. H. Poon. A new similarity measurement method for fuzzy-attribute graph matching and its application to handwritten character recognition. Fuzzy Logic Technology and Applications, Ed. By R. J. Marks, IEEE, Inc, New York City, 1994.

3.H.-C. Huang and S.-G. Miaou. A simplified neocognitron for recognition of handwritten Arabic numerals. 5th OCR & DA, pp 132-139, Taiwan, 1996.

4.J. Cho, M. Ahmadi, and M. Shridhar. A hierarchical neural network architecture for handwritten numeral recognition. Pattern Recognition, 30(2):289-294, 1997..

6.Ernst Kussul,Tatyana Baidyk,Improve Method of Handwritten Digit Recognition 2000