BL-YOLO-Seg: a lightweight instance segmentation method for metallic waste sorting
摘要
Efficient metallic waste sorting is important for resource recycling, low-carbon manufacturing, and sustainable industrial operation. However, practical visual perception in intelligent recycling systems remains challenging because metallic waste often exhibits reflective surfaces, irregular contours, large-scale variations, and cluttered backgrounds, which degrade segmentation accuracy and reduce the reliability of downstream robotic grasping and automated sorting. At the same time, resource-constrained edge environments and high-throughput recycling scenarios require lightweight models with real-time inference capability, making deployment-oriented visual intelligence an important issue for scalable industrial computing. To address these challenges, this study constructs the Metal Waste Instance Segmentation Dataset (MWISD) and develops BL-YOLO-Seg, a lightweight instance segmentation framework based on YOLOv11-Seg for metallic waste perception. The framework improves feature extraction, feature interaction, segmentation quality, and localization robustness through coordinated optimization of ADown, HyperC2Net, the LSCQ-Seg head, and SIoU, and is further optimized by pruning and knowledge distillation for deployment-oriented applications. Experiments on MWISD and two public datasets show that BL-YOLO-Seg achieves a favorable balance between accuracy and efficiency. On MWISD, the pruned and distilled BL-YOLO-Seg reduces parameters by 71.8% and computation by 39.4% compared with YOLOv11n-Seg, while improving F1(Box), mAP