基于SA-YOLOv8n的果园鸟类检测算法研究

Research on improved YOLOv8n-based avian detection in orchard

针对果园中鸟类检测模型参数量大、小目标检测能力不强,以及目标框回归的准确性和鲁棒性不足等问题,提出了一种基于YOLOv8n优化、改进的SA-YOLOv8n果园鸟类检测模型。模型采用平滑、连续可导的Mish激活函数替换SiLU激活函数,在Backbone部分添加ShuffleAttention模块,减少了模型大小。在Neck部分添加了第四个输出层,用于检测4×4以上的目标,优化了小目标检测。采用SIOU边界框回归损失函数替代CIOU,进一步提高了目标框回归的准确性和鲁棒性。实验证明,改进后的SA-YOLOv8n模型在自制鸟类数据集上的平均精度(AP)达到了96.40%,而单张图片检测仅需0.7 ms。与原YOLOv8n模型相比,改进后的模型在保持检测速度稳定的前提下,AP提高了1.6个百分点,模型大小降低了0.21 MB。这一系列改进不仅提升了性能的同时,还对模型进行了轻量化处理。

Aiming at the problems of large number of parameters in the bird detection model in the orchard,poor detection ability of small targets,and insufficient accuracy and robustness of the target frame regression,an optimised and improved SA-YOLOv8n orchard bird detection model based on YOLOv8n is proposed.The model replaces the SiLU activation function with a smooth and continuously derivable Mish activation function,and adds the ShuffleAttention module in the Backbone part to reduce the model size.A fourth output layer is added to the Neck part for detecting targets above 4×4 to optimise small target detection.The SIOU bounding box regression loss function is used instead of CIOU to further improve the accuracy and robustness of target box regression.Experiments demonstrate that the improved SA-YOLOv8n model achieves an average precision(AP)of 96.40%on the homemade bird dataset,while single-image detection takes only 0.7 ms.Compared with the original YOLOv8n model,the improved model improves the AP by 1.6 percentage points and reduces the model size by 0.21 MB,while maintaining a stable detection speed.this series of improvements not only improve the performance,but also lighten the model.