Agri-YOLO: An Improved YOLOv8 Algorithm for Farmland Obstacles Detection

This study proposes an improved object detection model, Agri-YOLO, based on the YOLOv8n baseline model, addressing the issues of insufficient detection accuracy and high model deployment costs for various types and scales of obstacles in agricultural fields. Three core optimization strategies are implemented to enhance model performance: replacing traditional convolution with the Wavelet Transform Convolution module to improve multi-scale feature perception with only a slight increase in parameters; utilizing the Wise-IoU loss function to optimize bounding box regression, enhancing the localization accuracy of irregular obstacles and effectively improving the convergence speed and regression accuracy of the loss function; and integrating the Dynamic Upsample module to reduce computational load while ensuring detection accuracy of agricultural obstacles, thereby improving feature recovery accuracy. Experimental results demonstrate that Agri-YOLO significantly outperforms baseline algorithms such as Faster R-CNN, SSD, and YOLOv8n in terms of precision, recall, and mAP50 metrics, with improvements of 0.4, 1.7, and 1.2% in accuracy, recall, and mAP0.5, respectively, while also enhancing model robustness and stability. This research provides an efficient technical solution for detecting obstacles in agricultural fields.