A lightweight YOLO-WRCS model for peanut pod detection in seed-preserving peanut shelling robots

Low-damage seed-preserving peanut shelling robots require real-time recognition of peanut pod quality and orientation to guide sorting and shelling while reducing seed damage.We proposed YOLO-WRCS, a lightweight YOLOv12n-based detector integrating RDWConv, SimAM, CARAFE, and Wise IoU. A dataset of 4,201 images covering five peanut pod classes was used for training and testing. The model was evaluated through ablation experiments, comparison with mainstream detectors, embedded deployment, and cross-variety validation.YOLO-WRCS achieved Precision, Recall, mAP50, and mAP50-95 values of 87.7%, 87.6%, 93.8%, and 92.7%, improving YOLOv12n by 2.2, 5.5, 3.4, and 3.2 percentage points. FLOPs and model size were reduced by 19.0% and 16.4%. On a Jetson Orin Nano platform, the model achieved 93.46% average recognition accuracy at 15.8 FPS and maintained above 90% accuracy across three untrained peanut varieties. YOLO-WRCS improves detection accuracy, lightweight design, and robotic deployment, although head-tail confusion and cross-variety robustness require further improvement.