Exploring Lightweight Vegetable Detection and Classification Models Based on YOLOv5s

At present, most of the vegetable detection machines are mobile terminals or mobile phones. In order to improve the reasoning speed of mobile devices, this paper conducts a lightweight improvement exploration based on YOLOv5s. Firstly, a dataset of 2500 vegetable images is built. To explore a better model, this paper adopted the method of comparative experiments, referring to ShuffleNetv2 and MobileNetv3 in lightweight network models, and made some changes to the backbone network of YOLOv5s. Finally, the modified YOLOv5S-shuffle and YOLOv5S-mobile were trained separately with the original YOLOv5s to compare the performance and effect differences of the three models. The experimental results show that compared to the original YOLOv5s, the parameters of the modified YOLOv5S-shuffle and YOLOv5S-mobile models are reduced by 57.9% and 47.2%, respectively. The mAP@.5 values of YOLOv5s-shuffle and YOLOv5s-mobile reached 85.5% and 93.4%, respectively.