Publication

Abstract :

The introduction of transformer models, which utilize a self-attention mechanism within deep neural networks, represents a notable breakthrough in natural language processing. This advancement has spurred researchers to investigate its applicability in computer vision tasks. Transformer-based models have showcased remarkable performance compared to traditional convolutional and recurrent neural networks across a spectrum of visual tasks, highlighting their prowess in representation learning. This study aims to classify vision transformer models according to their task-specific functionalities and conducts an extensive evaluation to discern their advantages and limitations. Moreover, it proposes effective strategies for integrating transformers into real-world device-based applications. Investigating the use of vision transformers, inspired by the transformer model, on diverse benchmark datasets, this research focuses on their ability to identify broad-based visual characteristics and distant dependencies through self-attentional mechanisms. Through meticulous examination on the Oxford Flowers dataset, this research delves into various pre-training methods, model topologies, and hyperparameters to enhance recognition accuracy. The findings reveal that vision transformers not only surpass previous state-of-the-art techniques but also exhibit the capacity to discern subtle differences among flower species with remarkable 98.5% accuracy. Furthermore, this research work explores the interpretability of the vision transformer model, elucidating how it recognizes and integrates crucial visual characteristics during classification. By analyzing the model’ s attention maps, insights into its decision-making process are provided. In essence, this research underscores the potential of vision transformers for fine-grained image classification and contributes to the burgeoning field of research on their efficacy in computer vision tasks.