FES-YOLOv5s: A Lightweight Model for Agaricus Bisporus Detection

Agaricus bisporus grows in complex environments and suffers from adhesion and occlusion problems. In this study, we propose a lightweight recognition model for Agaricus bisporus— FES-YOLOv5s—based on YOLOv5s. Our aim was to quickly and accurately detect Agaricus bisporus specimens. First, a FasterNet lightweight network was used in the backbone layer to reduce the computation of the model. An ECA mechanism was then introduced to enhance the interaction between multiple channels and improve the detection accuracy. Finally, a Soft-NMS module was used to replace the NMS module in YOLOv5s to resolve the missed detection of adherent and occluded Agaricus bisporus specimens. The improved model was named FES-YOLOv5s; F, E, and S represent the FasterNet, ECA, and Soft-NMS features, respectively. The FES-YOLOv5s model increased the mAP 0.5:0.95 by 2.4% and the FPS by 19.4%. It decreased the computation by 42.7% compared with the YOLOv5s model. The results of a comparison test revealed that the FES-YOLOv5s model demonstrated advantages in detection accuracy and speed compared with other target detection models. The FES-YOLOv5s model was deployed on an Agaricus-bisporus-picking robot; the detection success rate was greater than 90%, indicating that the improved model could detect Agaricus bisporus quickly and accurately in complex environments.