CWG-YOLOv8: a Navel Orange Detection Model Based on Improved YOLOv8 in an Agricultural Environment

Abstract The performance of picking robots for fruit object detection is crucial in agricultural environments. However, most existing detection models struggle to perform well in agricultural settings due to problems in detection accuracy, computational resource consumption, and real-time processing. To address these challenges, we propose a navel orange detection model called CWG-YOLOv8 based on YOLOv8, which can achieve accurate detection of navel oranges in agricultural environments. Firstly, we introduce a Convolutional Block Attention Module (CBAM) to enhance the backbone network and improve the generalization ability of the model. Secondly, Wise-IoU (WIoU) v3 as the bounding box regression loss function is employed, and a wise gradient allocation strategy is incorporated to emphasize high-quality samples, thus enhancing the model’s localization capability. Finally, we design a lightweight GhostNet module that effectively integrates shallow and deep features to reduce computational cost and speed up detection. The experimental results show that the number of parameters and the number of floating-point operations (FLOPs) of our model are reduced by 42.35 and 36.59% compared with the original model. After optimization and parameter training, the mean average precision average (mAP50) and mAP50∼95 of the model reached 95.1 and 83.3%, respectively. Compared with other mainstream models, our model demonstrates significant advantages in terms of detection accuracy, speed, and lightweight design, which meets the requirements of high real-time navel orange detection in agricultural environments.