基于改进YOLO v8n的轻量化茶叶嫩芽检测模型

Lightweight Tea Bud Detection Model Based on Improved YOLO v8n

摘要:

在智能采收名优茶过程中，现有的目标检测算法存在对茶叶嫩芽检测精确率不高和推理速度较慢的问题，导致在边缘计算设备上部署困难，为此本文提出一种基于改进YOLO v8n的轻量化茶芽检测模型YOLO-RET。通过引入重新设计的轻量化特征提取模块（RGCSPELAN模块）和改进的多尺度特征融合金字塔结构（EMBSFPN结构），在大幅降低模型参数量的前提下，提升了对目标特征信息的提取与融合能力。引入新的损失函数Focaler-IoU，解决了样本分布不均衡问题，进一步提高了模型检测精度和鲁棒性。试验结果表明，与YOLO v5n、YOLO v8n、YOLO v8s和YOLO v10n相比，YOLO-RET精确率分别提高3.2、3.0、0.8、4.3个百分点，平均精度均值分别提高2.7、3.2、0.4、3.1个百分点。此外，YOLO-RET模型参数量和浮点运算量相较原YOLO v8n模型分别降低43%和2.2×10?。将算法模型移植到ATK-DLRK3568开发板上，并进行量化处理以优化部署，降低了对硬件资源的需求。YOLO-RET模型在保持高识别精确率的同时，提升了检测效率，为在边缘计算端部署实时目标检测提供了一种高效、准确的解决方案。

Abstract:

During the intelligent harvesting of premium tea, existing object detection algorithms face challenges of insufficient detection precision for tender tea shoots and slow inference speed, making deployment on edge computing devices difficult. To address these issues, YOLO-RET, a lightweight tea shoot detection model was proposed based on an improved YOLO v8n architecture. By introducing a redesigned lightweight feature extraction module (RGCSPELAN module) and an enhanced multi-scale feature fusion pyramid structure (EMBSFPN), the model significantly enhanced feature extraction and fusion capabilities while substantially reducing model parameters. Additionally, a novel loss function, Focaler-IoU, was incorporated to address sample distribution imbalance, further improving detection accuracy and robustness. Experimental results demonstrated that compared with YOLO v5n, YOLO v8n, YOLO v8s, and YOLO v10n, YOLO-RET achieved precision improvements of 3.2, 3.0, 0.8, and 4.3 percentage points respectively, with corresponding mAP@0.5 improvements of 2.7, 3.2, 0.4, and 3.1 percentage points. Furthermore, YOLO-RET reduced the parameter count by 43% and computational complexity by 2.2 ×10? compared with that of the original YOLO v8n. The algorithm was successfully deployed on an ATK-DLRK3568 development board with quantization optimization, which not only lowered hardware resource requirements but also maintained high recognition accuracy while enhancing inference efficiency and reducing the computational burden on edge devices. The research result can provide an efficient and accurate solution for real-time object detection deployment on edge computing platforms.