YOLO and RF-DETR architectures for early disease diagnosis in Solanum tuberosum L

Abstract

Three object detectors for post-harvest diagnosis in tubers were compared under a unified protocol: two YOLO-family architectures and a transformer-based detector (RF‑DETR). We used a six-class public dataset (five diseases plus healthy). Lesions were manually annotated with bounding boxes in Roboflow and augmented with geometric/photometric transforms. Annotations were converted to COCO for RF‑DETR. We enforced identical splits (train/val/test = 792/56/57), the same input size (640 px), and an identical training recipe (60 epochs with early stopping). Evaluation on the held-out test set reported precision, recall, mAP@0.50, and mAP@0.50:0.95. On test images, YOLOv12 achieved P = 78.72%, R = 96.86%, mAP@0.50 = 79.36%, mAP@0.50:0.95 = 73.26%; YOLOv26 reached P = 79.34%, R = 96.64%, mAP@0.50 = 78.48%, mAP@0.50:0.95 = 73.55 (experimental logs). RF‑DETR delivered P = 70.86%, R = 88.39%, mAP@0.50 = 85.61%, and mAP@0.50:0.95 = 82.64%. For high-sensitivity screening (high recall), the YOLO architecture is more convenient, while RF-DETR offers strict localization (superior mAP at 0.50–.95) and simplifies post-processing by eliminating the need for an NMS. A two‑stage pipeline (YOLO → RF‑DETR) balances speed with spatial accuracy. Future work includes per‑class score calibration and multimodal fusion with NIR/HSI for subepidermal symptoms.