RF-DETR-Large for Nine-Class PCB Surface Defect Detection: An Empirical Evaluation of RF-DETR-Large on DsPCBSD+ with Three-Stage Continuation Training

Abstract

We report the best observed project result on the DsPCBSD+ nine-class PCB surface defect detection benchmark using RF-DETR-Large. Starting from a COCO-pretrained RF-DETR-L checkpoint, we apply a three-stage continuation training recipe: an initial retained stage at 640 × 640 for global epochs 1–3, a continuation at 640 × 640 with a learning rate of 1 × 10−4 through global epoch 7, and a final continuation at 640 × 640 with the learning rate reduced to 5 × 10−5 through global epoch 10. The best observed checkpoint at global epoch 9 reaches an EMA mAP@0.5:0.95 of 0.5411, exceeding the project's YOLO26x 100-epoch baseline reference of 0.4896 by +0.0515 absolute and +10.5% relative. This is a single-lineage validation result; no independent test estimate or uncertainty interval is claimed. The retained latency artifact reports a median of 35.3 ms per image on a single RTX 5090 with fp16 optimisation. The cited Zenodo record currently contains the paper source and PDF only; the local training artifacts are not represented as a complete public release.