Fig 1.
Roadmap for retail product object detection using lightweight methods.
Fig 2.
Overall structure diagram of YOLOv8n-improved model for the retail product object detection.
Fig 3.
Structure diagram of DWConv.
Fig 4.
Network structure of CBAM module.
Fig 5.
Structure diagram of CSPHet bottleneck and CSPHet-CBAM.
Fig 6.
It is used in YOLOv8n backbone to replace Conv.
Fig 7.
Schematic of channel pruning.
Fig 8.
The distribution of product categories and their quantities in the Locount dataset.
Table 1.
Experimental configuration.
Fig 9.
The training curves of metrics in YOLOv8n-improved model.
Table 2.
Ablation experiment results on Locount dataset.(“D” denotes DWConv, “CC” denotes CSPHet-CBAM, and “A” denotes ADown).
Table 3.
Model parameters and performance metrics under different channel pruning rates.
Fig 10.
Changes in the number of channels in the YOLOv8n-improved model at a pruning rate of 0.5.
Fig 11.
The distribution of Gamma coefficients () of the BN layers before YOLOv8n-improved model’s sparse training.
Fig 12.
Precision-Recall curve of LSR-YOLO model.
Table 4.
Performance comparison of the improved model before and after pruning.
Table 5.
Comparison of different object detection algorithms on Locount dataset.
Fig 13.
Comparison of heatmaps on retail products for different object detection models.
Table 6.
Ablation experiment results on COCO2017 dataset.(“D” denotes DWConv, “CC” denotes CSPHet-CBAM, and “A” denotes ADown).
Table 7.
Comparison of detection performance between LSR-YOLO and other models on COCO2017 dataset.
Fig 14.
Comparison of heatmaps for different object detection models on COCO scenes.
Table 8.
The 5-fold cross validation result of LSR-YOLO on Locount dataset.
Fig 15.
Examples of negative samples for product detection with multiple object occlusions.
Fig 16.
Examples of negative samples for product detection under low-light conditions.