A lightweight and efficient carton detection model for intelligent warehouse environments

Abstract

Carton detection is a crucial task in intelligent warehouse systems, whose performance is primarily constrained by the limited storage and computing resources of warehouse edge devices. In this paper, a new model called BM-YOLO is proposed based on You Only Look Once 11 (YOLO11) to overcome these issues. First, a DimPool attention module is introduced into the backbone network, which integrates multidimensional collaborative attention with spatial pyramid pooling fast to enhance feature interactions and improve target discrimination. Second, a multiscale cooperative feature pyramid network is constructed in the neck network by combining the bidirectional feature pyramid network with multidimensional collaborative attention. This structure adaptively models cross-scale semantic relationships while reducing computational redundancy. Third, the original large detection head is removed to make the model lightweight without sacrificing detection accuracy. Experimental results in three public datasets validate the superiority of the BM-YOLO model compared with the baseline model YOLO11. In the live stacked carton dataset, BM-YOLO has achieved an increase of 33.2% in frames per second and an increase of 0.6% in precision of mAP@0.5. In the online stacked carton dataset, BM-YOLO achieves an improvement of 0.8% in precision along with an increase of 0.3% in both mAP@0.5 and mAP@0.5:0.95. In the logistics objects in context dataset, BM-YOLO achieves a 0.8% increase in precision and a 0.6% improvement in mAP@0.5, further validating the model’s generalization capability.