Efficient Image Retrieval With Transformer and Residual Networks Fusion

The growing use of images across various applications has highlighted the need of effective image search systems. Traditional search methods based on text frequently prove inadequate in this context, prompting the emergence of Content-Based Image Retrieval (CBIR) systems. While existing CBIR systems usually concentrate on extracting basic features such as texture, shape, and color, they frequently encounter challenges related to both accuracy and speed. To address these limitations, we propose a novel CBIR approach that integrates the Vision Transformer (ViT) and the Residual Networks (ResNet) architectures for feature extraction. This fusion allows us to extract more precise and meaningful features from images, thereby enhancing the overall effectiveness of our system. Additionally, we employed the Principal Component Analysis (PCA) to reduce the feature space's dimensionality, leading to improved efficiency. To facilitate rapid and efficient similarity searches, we implemented the Annoy library. The experimental results demonstrated that our proposed CBIR system surpasses existing systems in both accuracy and speed, underscoring the effectiveness of our approach in image retrieval tasks.