|
DATA AUGMENTATION AND NEURAL NETWORK MODELS CONSTRUCTION FOR HAND-WRITTEN CHARACTER RECOGNITION IN BIOMETRIC AUTHENTICATION SYSTEMS |
|---|---|
| รหัสดีโอไอ | |
| Creator | Anastasia Gennadevna Isaeva, Alexey Sergeevich Katasev, Amir Muratovich Akhmetvaleev, Dina Vladimirovna Kataseva |
| Title | DATA AUGMENTATION AND NEURAL NETWORK MODELS CONSTRUCTION FOR HAND-WRITTEN CHARACTER RECOGNITION IN BIOMETRIC AUTHENTICATION SYSTEMS |
| Contributor | - |
| Publisher | TuEngr Group |
| Publication Year | 2562 |
| Journal Title | International Transaction Journal of Engineering, Management, & Applied Sciences & Technologies |
| Journal Vol. | 10 |
| Journal No. | 13 |
| Page no. | 10A13C: 1-8 |
| Keyword | biometric identification, Augmented ANN, Distorted handwriting, Numeric handwriting, Training ANN. |
| URL Website | http://tuengr.com/Vol10_13.html |
| Website title | ITJEMAST V10(13) 2019 @ TuEngr.com |
| ISSN | 2228-9860 |
| Abstract | This paper solves the problem of neural network model construction for hand-written character recognition. The preparation of data for training neural networks was carried out using the augmentation method by including distortions of two types in the data: resizing and rotation of the characters. To build the models, two training sets were used: initial data and augmented data. The training accuracy of neural network models was 100%. To assess the generalization ability of the constructed models, a test sample with a volume of 1000 records (100 records for each digit) was formed, consisting of ordinary and distorted images of decimal digits missing in the training samples. The accuracy of the initial neural network model during testing was 84.7%, and the accuracy of the augmented model was 95 %. For each neural network model, errors of types I and II were calculated, having a value of 6.2% and 9.1% for the original neural network model and 1.8% and 3.2% for the augmented one. Consequently, the enrichment of the initial data allowed us to reduce the level of errors of the first type by 4.4%, and the number of errors of the second type by 5.9%. In addition, the total error of each model is calculated based on the three-block cross-validation method. The error of the original model was 0.21, and the augmented one - 0.07. Consequently, the total error of the augmented model decreased by 0.14 compared with the original model. Thus, the augmented neural network model is an effective hand-written character recognition tool and can be used in biometric authentication systems. |
