T2T-ViT: A Novel Semantic Image Mining Approach for Improving CBIR Using Vision Transformer

Abstract In the field of Image Mining (IM) and Content-Based Image Retrieval (CBIR), the significance lies in extracting meaningful information from visual data. By focusing on the intrinsic meaning within images, semantic features enhance the accuracy and relevance of image retrieval systems, bridging the gap between human understanding and computational analysis in visual data exploration. This research explores the fusion of image processing techniques and CBIR. The need for this research is based on the persistent challenges in existing CBIR systems, where traditional methods often fall short of comprehensively capturing the intricate semantics of images. The primary objective of this research is to propose a novel approach to CBIR by implementing the Tokens-to-Token Vision Transformer (T2T-ViT) to address the limitations of traditional CBIR systems and enhance the accuracy and relevance of image retrieval. The T2T-ViT model achieves exceptional performance in CBIR on Corel datasets, with a high accuracy of 99.42%, precision of 98.66%, recall of 98.89%, and F-measure of 99.35%. The model demonstrates a harmonious balance between identifying and retrieving relevant images compared to existing models.