基于YOLO v5n-DRSW的机采茶青芽叶形态智能检测研究

Intelligent Morphological Detection Research of Mechanically Picked Tea Green Buds and Leaves Based on YOLO v5n-DRSW

摘要:

精准检测机械采收茶青形态,有利于提高机采茶青自动化分级精度及效率。本研究构建了YOLO v5n-DRSW机采茶青形态检测模型,以YOLOv5n为基线,通过引入分布移位卷积模块(Distribution shifting convolution,DSConv)至头部网络,在减小了模型复杂度的同时提高了推理速度。此外,在颈部网络引入重参数化泛化特征金字塔网络结构(Reparameterized generalized-FPN,RepGFPN),提升了模型的泛化能力和鲁棒性。针对茶青嫩芽小目标特征,将压缩和激励注意力机制(Squeeze and excitation,SE)嵌入骨干网络,借助全局视野增强对特征图的感知能力,进而提高检测精度。最后,将加权交并比损失函数(Wise intersection over union,WIoU)融入网络的损失计算部分,以动态调整梯度增益。试验结果表明,相较于基线模型,YOLO v5n-DRSW在检测复杂形态机采茶青方面表现出显著优势:模型精确率提升了2.1个百分点,达98.1%;在推理速度方面,YOLOv5nDRSW处理单帧图像仅需2.11ms,较基线模型有显著提升;YOLO v5n-DRSW的浮点运算次数相较于基线模型减少了2.44%,有效保持了轻量化。实际应用结果表明,机采茶青形态在线检测平均准确率达94.34%,漏检率不超过1.1%,检测性能较好。

Abstract:

Accurate morphological detection of mechanically harvested fresh tea leaves is beneficial for improving the accuracy and efficiency of automated grading. YOLO v5n-DRSW, an advanced YOLO v5n based model specifically designed for the precise recognition of machine-harvested tea leaf morphologies was introduced. The model integrated several innovative features: a distribution shifting convolution (DSConv) module in the head network to reduce complexity and enhance efficiency; a reparameterized generalized feature pyramid network (RepGFPN) in the neck to improve generalization and robustness; and the squeeze-and-excitation (SE) attention mechanism embedded in the backbone to strengthen feature perception for small targets like tender buds. By leveraging a global field of view, this mechanism further enhanced the perception capability of the feature maps. Additionally, the wise intersection over union (WIoU) loss function was used to dynamically adjust gradient contributions during training. Compared with the baseline, YOLO v5n-DRSW exhibited significant advantages in detecting machine-harvested tea leaves with complex morphologies. Experimental results demonstrated that YOLO v5n-DRSW achieved 98. 1% accuracy, a 2. 1 percentage points improvement over the baseline, with an inference time of just 2. 11 ms per frame. This rapid processing speed represented a notable improvement over the baseline model. The model also reduced floating-point operations by 2. 44% , confirming its lightweight nature. In practical applications, it attained an average online recognition accuracy of 94. 34% with a miss rate below 1. 1% , highlighting its strong potential for enhancing automated tea leaf grading systems. Overall, the model demonstrated excellent detection performance, providing reliable assistance for improving the morphological detection outcomes of fresh tea leaves.