Research on Lightweight Model for Visual Detection of Tomato Maturity in Natural Scenes

Conducting maturity detection for tomato in a natural scenes is a prerequisite for achieving automated tomato harvesting. This study aims to conduct research on a lightweight visual detection model for tomato maturity, which is deployed on the edge of tomato picking robots. Firstly, based on the natural scenes of tomato plantations, taking pictures from the viewpoint of a robot to create tomato maturity detection dataset. Then, based on the YOLO v5s(You Only Look Once version 5 Small) , an improved model named YOLOv5s BCD is proposed, which integrates the CA attention mechanism, BiFPN neck feature fusion network, and Decoupled Head, this model serves as the tomato maturity detection model in this study. Finally, experimental research on model training and testing was conducted, and the model was evaluated and refined through model ablation experiments and comparative experiments. The experimental results indicate that the proposed improved model, YOLOv5s BCD, in this study has a memory usage of 25.6MB, with an average precision mean (mAP) of 95.6% and an inference speed of 28.5FPS, Its comprehensive performance is better than other models in the YOLO v5 series, as well as YOLO v7tiny, YOLO v8 and YOLO v9c. It is suitable for lightweight and real-time deployment on the edge device, and can serve as a visual system for tomato picking robots, providing technical support for automated picking.