Sharif Digital Repository

In addressing the challenge of image similarity estimation on the MNIST dataset, our research drives from conventional Siamese network methodologies by incorporating Vision Transformer (ViT) architecture. Departing from the standard MNIST dataset, we introduced a novel paired dataset tailored to enhance the capabilities of similarity estimation. The innovation lies in the utilization of ViT as the core foundation for extracting features, followed by the application of Euclidean distance metrics on the dual input. This departure from the traditional approach not only broadens the scope of image similarity assessment but also enhances the model's discriminative power. Notably, the model... 

This study focuses on the task of Persian numeral classification within image data, employing the Vision Transformer (ViT) architecture to predict numerals akin to the MNIST dataset, but adapted to the Persian script. Our approach yielded a notable validation accuracy of 0.9920, particularly noteworthy when employing a patch size of 4. Notably, the research introduces an innovative visualization aspect, showcasing the first multi-head attention linear map and its counterpart, the last one. The visualization of these attention maps provides a unique insight into the model's internal processes and highlights its proficiency in capturing intricate patterns within Persian numeral images. This...