A Lightweight YOLOv8-Based Model for Tomato Objection and Keypoint Detection

In response to the challenge of balancing accuracy and real-time performance in target detection and keypoint detection for tomato cluster harvesting automation, this study proposes an improved lightweight and high-precision YOLOv8 multi-task detection model. By integrating a lightweight wavelet convolution module, optimizing the neck network for cross-scale feature fusion, and replacing the upsampling module, the model significantly reduces parameter size and computational cost while enhancing the detection accuracy of tomato clusters in complex environments. Additionally, the loss function is improved by introducing auxiliary boxes and a dynamic focusing mechanism, which helps detect densely packed small targets in complex environments. Experimental results show that the improved model achieves 93.4% accuracy in object detection and 92.9% accuracy in keypoint detection on a self-built tomato bunch dataset. The model's parameter size is reduced by 36.4%, and Gflops decreased from 8.3 to 6.6. Both object detection and keypoint detection accuracy have increased by 3.3%. Compared to the latest YOLOvll model, the improved model has 26% fewer parameters, with similar performance in both object detection and keypoint detection. This improved model is suitable for resource-constrained scenarios such as agricultural harvesting robots and provides an efficient and reliable technical solution for fruit and vegetable harvesting automation.