基于YOLO v8_EGW的马铃薯种薯芽眼检测方法

Detecting Method of Potato Seed Bud Eye Based on YOLO v8_EGW

摘要:

针对传统目标检测算法在马铃薯种薯芽眼检测中易受泥土覆盖、表面破损以及环境因素等影响，造成检测效果差的问题，提出了一种YOLO v8_EGW模型，在自制的马铃薯数据采集试验台上实现了芽眼的快速准确检测。在YOLO v8的主干部分引入EMA注意力机制来提升马铃薯表面的芽眼特征提取能力；在YOLO v8的颈部网络中引入GD机制，并使其与C2F模块结合，加强对芽眼特征的信息融合能力，提升芽眼特征的检测能力；替换位置损失函数为WIoU损失函数，提高标记框的质量并进一步提升检测性能。结果表明，改进模型精确率、召回率、mAP@0.5、mAP@0.5:0.95分别为95.8%、92.7%、94.8%、73.0%，模型内存占用量为40.4MB、检测速度为37.03f/s。与同类目标检测模型（YOLO v4、YOLO v5、YOLO v6、YOLO v7、YOLO v8n）相比，其检测精确率分别高4.5、3.1、5.2、4.7、4.1个百分点，检测结果均优于其他模型。研究结果可为后续智能切种机的研发提供理论支持。

Abstract:

In order to solve the problem that the traditional object detection algorithm is susceptible to the influence of soil covering, surface damage and environmental factors in potato sprout eye detection, resulting in poor detection effect, an YOLO v8_EGW model was proposed, which realized the fast and accurate detection of potato sprout eyes on the self-made potato data acquisition experimental bench. Firstly, in order to improve the ability to extract bud eye features on the potato surface, the EMA attention mechanism was introduced into the backbone part of YOLO v8. Secondly, the GD mechanism was introduced into the neck network of YOLO v8 and combined with the C2F module to strengthen the information fusion ability of bud eye features and improve the detection ability of bud eye features. Finally, the position loss function was replaced with the WIoU loss function to improve the quality of the marking box and further enhance the detection performancee. The results showed that the precision, recall, mAP@0.5, mAP@0.5:0.95 of the improved eye detection model were 95.8%, 92.7%, 94.8% and 73.0%, respectively. The model size was 40.4MB and operated at a detection speed of 37.03 frames per second. Compared with similar object detection models of YOLO v4, YOLO v5, YOLO v6, YOLO v7 and YOLO v8n, the detection accuracy was 4.5, 3.1, 5.2, 4.7 and 4.1 percentage points higher, respectively, and the detection effect was better than other models. Moreover, it exhibited lower rates of false detections (3.7%) and missed detections (1.5%) compared with other models. These enhancements ensured the model’s capability to detect potato seed bud eyes effectively under diverse conditions, offering theoretical support for the development of intelligent seed cutting machines, which could revolotionize agricultural automation in the near future.