Comparing Six Deep Learning Feature Extractors for CBIR Tasks

This article evaluates six deep-learning feature extractors for content-based image retrieval (CBIR), spanning both self-supervised and supervised approaches. It analyzes DINOv1, DINOv2, and DreamSim as ImageNet-pretrained self-supervised models, and contrasts them with SwinTransformer and two ResNet50 variants—one trained on RadImageNet and another on fractal geometry renderings. By extending earlier studies, the comparison highlights how backbone choice, training data, and pretraining strategies impact performance across medical and synthetic imaging tasks.