YOLO-LDB: A YOLOv11-based lightweight dual-backbone method for dual-energy X-ray ore detection

Efficient and accurate ore detection critically impacts overall productivity in mineral processing automation, but high model complexity, excessive computational resource demands, and limited adaptability to complex ore characteristics hinder existing object detection algorithms. To address the limitations of existing object detection algorithms in terms of model complexity, computational resource consumption, and handling com-plex ore images. This study integrates dual-energy X-ray imaging technology with advanced object detection to propose a lightweight dual-backbone YOLO (YOLO-LDB) model based on YOLOv11 architecture, achieving highly efficient and accurate ore grade detection. The optimized network structure employs pruning techniques while introducing an auxiliary backbone to compensate for potential accuracy loss. By replacing conventional convolution with modulated deformable convolution, the model enhances adaptability to diverse ore morphology. A multi-scale feature fusion module facilitates inter-backbone feature transmission, improving detection performance across varying scales. The lightweight detection head adopts parameter sharing strategy, and the use of SIoU loss contributes to more precise bounding box regression. The experimental findings indicate that the model achieves excellent results, including a mAP@0.5:0.95 of 0.977, detection speed of 69.9 FPS, computational cost of 1.5G FLOPs, 2.57M parameters, and a lightweight size of 1.63MB.