A lightweight YOLOv11-based commodity detector for dense smart retail cabinets

The development of smart retail cabinets is tightly intertwined with machine vision technology, yet most confront the bottleneck of limited commodity variety. This is primarily attributed to the vast diversity of commodities, which gives rise to multiple recognition challenges–including small target detection, occlusion, and particularly the identification of similar items. To address these critical engineering issues, this paper proposes an improved lightweight YOLO11 commodity recognition algorithm based on a YOLOv11-based lightweight detection framework with structural refinement and stabilized IoU regression. First, the original C3K2 module is upgraded to C3K2-MKFENet, enhancing feature extraction via partial convolution and multi-scale strategies. Second, the CARAFE module replaces the original sampling component, and a fuseModule is introduced to fuse high-resolution shallow features with deep semantic features, strengthening small target fitting. Third, the CBAM module is integrated to prioritize key features for occlusion mitigation. The XIoU loss is also introduced to improve CIoU efficiency. Experiments on the YZGOODS dataset show mAP50 and mAP50-95 increase by 0.6% and 2.3% respectively; we report GPU reference speed under a unified measurement protocol, and we additionally report an embedded-device benchmark on Jetson Nano, showing a practical accuracy–efficiency trade-off under resource constraints.