ANALYSIS OF NEURAL NETWORK ARCHITECTURES FOR RECOGNITION OF HANDWRITTEN NUMBERS

DOI: 10.31673/2412-4338.2023.041827

Abstract

The accuracy of contemporary systems designed for the recognition of handwritten numbers exhibits significant variability, contingent upon the technologies selected during the development of a specific method. Neural networks have supplanted algorithms reliant on statistical data and datasets for this task, primarily due to their better performance.
Several recent publications addressing the challenges of handwritten digit recognition employing diverse types of neural networks have been examined. This paper delves into the analysis of the architecture of two prominent types: dense neural networks and convolutional neural networks along with their respective training methods applicable to the recognition of handwritten digits from images.
The study's findings reveal that Dense Neural Networks (DNN), while applicable to the task at hand, demonstrate diminished performance as the size of the input image increases. Conversely, Convolutional Neural Networks (CNN) prove more adept at image analysis, owing to their convolution and pooling operations, thus establishing their priority in this context.
Four distinct types of learning applicable to neural networks were explored, namely Supervised Learning, Unsupervised Learning, Semi-supervised Learning, and Reinforcement Learning. The investigation concludes that Supervised Learning using backpropagation emerges as the most fitting method for training neural networks to address the identified problem. The paper further explores the stages and principles underpinning the operation of the backpropagation method.
Building upon the derived results, there is a plan to develop an application leveraging the identified architecture. This application aims to empower users with the capability to efficiently extract digital information from images, irrespective of their resolution. The envisioned outcome is an application that attains high accuracy and processing speed in image recognition and extraction.

Keywords: neural networks, handwritten digits, convolutional neural networks, dense neural networks, neural network learning algorithms, backpropagation.

References:
1. Comprehensive review of artificial neural network applications to pattern recognition [Electronic resource] / Oludare Isaac Abiodun [et al.] // IEEE access. – 2019. – Vol. 7. – P. 158820–158846. – Access mode: https://doi.org/10.1109/access.2019.2945545 (access date: 04.12.2023).
2. A survey on using neural network based algorithms for hand written digit recognition [Electronic resource] / Muhammad Ramzan [et al.] // International journal of advanced computer science and applications. – 2018. – Vol. 9, No. 9. – Access mode: https://doi.org/10.14569/ijacsa.2018.090965 (access date: 03.12.2023)
3. Al-Taee M. M. Handwritten Recognition: a survey [Electronic resource] / May Mowaffaq Al-Taee, Sonia Ben Hassen Neji, Mondher Frikha // 2020 IEEE 4th international conference on image processing, applications and systems (IPAS), Genova, Italy , December 9–11 2020 - [B. m.], 2020. – Access mode: https://doi.org/10.1109/ipas50080.2020.9334936 (access date: 12/01/2023)
4. MNIST handwritten digit recognition using machine learning [Electronic resource] / Elizabeth Rani G [et al.] // 2022 2nd international conference on advance computing and innovative technologies in engineering (ICACITE), Greater Noida, India, April 28–29. 2022 - [B. m.], 2022. – Access mode: https://doi.org/10.1109/icacite53722.2022.9823806 (access date: 12/03/2023)
5. Digit recognition of MNIST handwritten using convolutional neural networks (CNN) [Electronic resource] / Sakthimohan M [et al.] // 2023 international conference on intelligent systems for communication, iot and security (iciscois), Coimbatore, India, 9– February 11 2023 - [B. m.], 2023. – Access mode: https://doi.org/10.1109/iciscois56541.2023.10100602 (access date: 12/03/2023)
6. Simonyan K. Very Deep Convolutional Networks for Large-Scale Image Recognition / Karen Simonyan, Andrew Zisserman // ICLR 2015, San Diego, July 7. 2014 - [B. m.].
7. What is UNZR and where to get it? [Electronic resource] // State Migration Service. – Access mode: https://dmsu.gov.ua/news/region/6563.html (date of application: 11.11.2023)
8. Dumoulin V. A guide to convolution arithmetic for deep learning / Vincent Dumoulin, Francesco Visin. - [B. m.] : Monrea. University, 2018. - 31 p.
9. LeCun Y. Deep learning [Electronic resource] / Yann LeCun, Yoshua Bengio, Geoffrey Hinton // Nature. – 2015. – Vol. 521, No. 7553. – P. 436–444. – Access mode: https://doi.org/10.1038/nature14539 (access date: 05.12.2023)
10. Improved handwritten digit recognition using convolutional neural networks (CNN) [Electronic resource] / Savita Ahlawat [et al.] // Sensors. – 2020. – Vol. 20, No. 12. – P. 3344. – Access mode: https://doi.org/10.3390/s20123344 (access date: 02.12.2023)
11. Zhang D. The application research of neural network and BP algorithm in stock price pattern classification and prediction [Electronic resource] / Dehua Zhang, Sha Lou // Future generation computer systems. – 2021. – Vol. 115. – P. 872–879. – Mode of access: https://doi.org/10.1016/j.future.2020.10.009 (access date: 05.12.2023)