Lightweight optimization of YOLOv8m for robotic vision-based snack cucumber sorting and palletizing

Abstract

To address the limited research on snack cucumber plate-setting robots and the difficulty of deploying advanced deep learning models on edge devices with constrained hardware resources, this paper proposes a lightweight improvement of the YOLOv8m model for this application scenario. First, the FasterNet Block replaces the Bottleneck structure in the original C2f module, effectively reducing computational redundancy and memory access frequency while maintaining high detection accuracy, thereby enabling more efficient spatial feature extraction. Second, the AFPN module is introduced to replace the original small-object detection head, further enhancing the model's ability to detect small targets and improving overall performance. Finally, the WIoU v3 loss function is adopted in place of the original CIoU to improve localization accuracy and convergence speed. Experimental results show that, compared with the original YOLOv8m model, the improved version reduces the number of parameters, model size, and floating-point operations by 63.15 %, 62.02 %, and 39.44 %, respectively, while increasing detection accuracy by 0.2 percentage points. The proposed model achieves a lightweight design without sacrificing detection performance, offering effective technical support for cucumber plate-setting robots on resource-constrained hardware, and potentially informing future efforts in the development of intelligent agricultural equipment. • This study proposes a lightweight recognition approach for the snack cucumber sorting and stacking robot. • The Bottleneck module in the C2f structure was replaced with the FasterNet Block to reduce computational redundancy and frequent memory access, thereby achieving model lightweighting. • To further reduce model complexity, the small object detection head was replaced with the AFPN, enhancing lightweight performance. • The original CIoU loss in the network was replaced with WIoU v3 to improve detection accuracy. • The proposed method demonstrated favorable performance in validation.