2017³â ÄÄÇ»ÅÍÁ¾ÇÕÇмú´ëȸ
Current Result Document : 3 / 3
ÇѱÛÁ¦¸ñ(Korean Title) |
Comparison of Neural Network Methods for Handwriting Datasets |
¿µ¹®Á¦¸ñ(English Title) |
Comparison of Neural Network Methods for Handwriting Datasets |
ÀúÀÚ(Author) |
Athita Onuean
Ryong Lee
Jangwon Gim
Taehon
|
¿ø¹®¼ö·Ïó(Citation) |
VOL 44 NO. 01 PP. 0893 ~ 0895 (2017. 06) |
Çѱ۳»¿ë (Korean Abstract) |
Nowadays, Deep Learning models have been applied to fields such as computer vision, image recognition and handwritten character recognition as a promising method due to the advance of software and hardware environments. In this paper, we tried to develop a recognizer for handwritten digits by Thai students. We applied three representative learning models (Logistics,
MLP, and CNN) to train with our datasets (THD: Thai Handwritten Data) and compared the accuracy rate with a public dataset called MNIST. In the comparison among the learning models in term of an accuracy rate with our datasets, we found that a small scale of training dataset makes it difficult to obtain high accuracy results. In order to tackle this problem, we adopted the Fine-Turning with Transfer Learning strategy to take advantages of already trained models. The result by the Fine-Turning technique in the training set have been yielded more height accuracy rate when compared with using THD as a training set and combined data on MNIST+THD as a training set. From the results, the accuracy rate of CNN: LeNet-5 increased up to 2.6% when using combination of MNIST and THD, whereas MLP methods (ReLu and sigmoid) increased up to only 5%.
|
¿µ¹®³»¿ë (English Abstract) |
Nowadays, Deep Learning models have been applied to fields such as computer vision, image recognition and handwritten character recognition as a promising method due to the advance of software and hardware environments. In this paper, we tried to develop a recognizer for handwritten digits by Thai students. We applied three representative learning models (Logistics,
MLP, and CNN) to train with our datasets (THD: Thai Handwritten Data) and compared the accuracy rate with a public dataset called MNIST. In the comparison among the learning models in term of an accuracy rate with our datasets, we found that a small scale of training dataset makes it difficult to obtain high accuracy results. In order to tackle this problem, we adopted the Fine-Turning with Transfer Learning strategy to take advantages of already trained models. The result by the Fine-Turning technique in the training set have been yielded more height accuracy rate when compared with using THD as a training set and combined data on MNIST THD as a training set. From the results, the accuracy rate of CNN: LeNet-5 increased up to 2.6% when using combination of MNIST and THD, whereas MLP methods (ReLu and sigmoid) increased up to only 5%.
|
Å°¿öµå(Keyword) |
|
ÆÄÀÏ÷ºÎ |
PDF ´Ù¿î·Îµå
|