SSL-YOLO

SSL-YOLO is a research framework that addresses the challenge of training object detectors with very few labeled samples. It combines contrastive self-supervised pre-training (SimCLR) with YOLOv8 fine-tuning, enabling better detection performance from as few as 10 labeled images per class. The pipeline extracts the YOLOv8 backbone, pre-trains it on unlabeled industrial images, then transfers learned representations to YOLOv8 for downstream defect detection. All experiments are implemented in PyTorch and inference with Ultralytics.