Enhanced Recognition: A Vision Transformer Approach for Digits

Vision Transformers (ViTs) are good at applying self-attention mechanisms to learn global information, but require large datasets, are noisy, and are computationally expensive. With the introduction of few-shot learning, which reduces complexity and improves detection, particularly on small datasets like MNIST, our study overcomes these limitations. This flexibility ensures stable performance with a small number of labeled data while reducing reliance on large datasets. Our optimized model is highly effective for real-time applications, achieving an accuracy rate of 94% and an inference time of just 0.8 milliseconds per image. This study introduces a new, scalable approach to handwritten digit recognition in resource-limited environments by shifting from the feature-rich extraction ability of ViTs to the flexibility provided by few-shot learning, thus reducing computational expense and vulnerability to noise.