Related Experiment Video
Updated: Jul 4, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.8K
Performance analysis of hybrid deep learning framework using a vision transformer and convolutional neural network
Vanita Agrawal1, Jayant Jagtap2, Shruti Patil3
1Department of Computer Science and Information Technology, Symbiosis Institute of Technology, Symbiosis International (Deemed University), Pune, Maharashtra, India.
Methodsx
|January 31, 2024
Summary
This study introduces a hybrid convolutional Vision Transformer (ViT) model for improved handwritten digit recognition. The model demonstrates superior accuracy on both clean and noisy real-world data.
Area of Science:
- Computer Science
- Artificial Intelligence
- Machine Learning
Background:
- Digitization necessitates efficient handwritten document recognition systems, with accurate digit recognition being crucial for data analysis.
- Convolutional Neural Networks (CNNs) achieve high accuracy but lack adaptability due to fixed filter weights.
- Vision Transformers (ViTs) and Multilayer Perceptrons (MLPs) offer dynamic processing, leading to interest in hybrid models.
Purpose of the Study:
- To analyze the effectiveness of a hybrid convolutional Vision Transformer (ViT) model for handwritten digit recognition.
- To evaluate the model's performance on both cleaned and uncleaned (real-time) handwritten digit images, considering noise and distortions.
- To compare the proposed model's accuracy against state-of-the-art techniques.
Main Methods:
- Developed and evaluated a hybrid model combining convolutional neural network (CNN) features with Vision Transformer (ViT) architecture.
- Utilized both cleaned and uncleaned handwritten digit datasets to simulate real-world conditions.
- Implemented cross-validation and hyper-parameter tuning to optimize model performance.
Main Results:
- The hybrid convolutional ViT model achieved the highest recognition accuracy compared to existing methods.
- The model demonstrated robustness and effectiveness on both clean and noisy handwritten digit images.
- Performance enhancements were observed through cross-validation and meticulous hyper-parameter tuning.
Conclusions:
- The proposed hybrid convolutional ViT model offers a robust, feasible, and highly accurate solution for handwritten digit recognition.
- The findings suggest potential for extending this approach to recognize other elements within handwritten documents.
- Hybrid models represent a significant advancement in overcoming the limitations of traditional CNNs for dynamic pattern recognition.
