YOLO-FCAP: An improved lightweight object detection model based on YOLOv8n for citrus yield prediction in complex environments

In citrus orchard, accurate object detection serves as a critical technology and an essential prerequisite for achieving precise yield prediction. However, the complex orchard environments where fruit and leaves obscure each other and the limited capabilities of edge devices pose challenges for the implementation of high-precision, lightweight detection algorithms. Using object detection algorithms to model data based on predicted and actual yields, it is feasible to predict the overall production of citrus trees in orchards effectively. To address the aforementioned challenges for efficient and accurate yield prediction, this study proposed YOLO-FCAP, a lightweight model based on an improved version of YOLOv8n. Firstly, the smaller network scaling factor called pico was introduced to reduce network redundancy and achieve a significant decrease in model size while ensuring performance. Secondly, all of the C2F modules were replaced with the FasterNetBlockWithCA module, which possesses stronger anti-background interference capability and a lighter-weight structure. Thirdly, the last three downsampling convolutional blocks of the backbone network were replaced with the ADown module to enable multi-scale feature fusion and improve the model's ability to retain detailed information. Finally, the small object detection layer was added to improve the model's accuracy for detecting smaller and occluded targets. The experimental results obtained on the self-constructed citrus dataset showed that YOLO-FCAP's parameters and FLOPs were reduced to 0.81M and 5.5G, respectively. Its precision, recall, and average precision reached 90.4%, 84.3%, and 92.5% respectively, while the frame rate achieved 168.42 frames per second. In addition, this study obtained a linear regression relationship between predicted and actual yields based on some of the samples, where the coefficient of determination R 2 reached 0.98, with an RMSE of 1.05 and an MAE of 0.83, indicating a strong correlation between predicted and actual yields, with minimal estimation error. The results demonstrated that YOLO-FCAP can efficiently predict yield in complex orchard environments, which provides technical support for orchard planning at various stages.