A Lightweight YOLO Model for Commodity Recognition

This paper presents EffiRepYOLO, a novel lightweight model optimized for real-time commodity recognition in resource-constrained environments. By integrating EfficientNetV2 and RepVit into the YOLOv8 framework, the proposed model achieves a superior balance between accuracy and computational efficiency. The redesigned backbone replaces CSPDarknet53 with EfficientNetV2's FusedMBConv, and MBConv modules, reducing parameters by 15 % while maintaining feature extraction capability. Additionally, RepVit's reparameterized multi-head self-attention mechanism enhances global contextual understanding in the Neck section. Evaluated on the MS COCO dataset, EffiRepYOLO achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 5. 1 \%} \text{mAP} {@} 0.5$</tex> and 54.3% [email protected]:0.95 with only 2.7M parameters and 8.2G FLOPs, outperforming existing lightweight models like YOLOv7-tiny and YOLOv8 variants. The proposed model provides a practical solution for deploying high-accuracy commodity recognition systems on resource-constrained edge devices.