Fine-tuning vision transformers for bacterial image classification using the DIBAS dataset

Since the emergence of Vision Transformer (ViT), this model and its derivative architectures have achieved remarkable performance in various computer vision tasks. However, their extensive computational requirements and dependence on massive datasets pose challenges for bacterial classification tasks with limited data. This study investigates the effectiveness of employing transfer learning with ViT on the DIBAS bacterial image dataset and compares its performance with that of conventional convolution-based models. We explore various transfer learning strategies, analyze the influence of patch size, pretraining dataset, and model depth, and present our experimental findings. Detailed explanations of the key methodologies are also provided to enhance understanding.