基于改进YOLO v5n的移栽机栽植部件辣椒苗识别方法

Pepper Seedling Recognition in Transplanting Machine Components Based on Improved YOLO v5n

摘要:

全自动移栽机在实际工作过程中经常遇到堵苗、漏苗和幼苗栽植状态异常等问题，实时监测移栽机栽植部件中的幼苗情况，是提高移栽机工作效率与移栽质量的关键。因此，本文提出了一种基于YOLO v5n的轻量化识别方法，用于对栽植部件中的辣椒苗进行精确目标检测。首先在顺光、逆光光照情况下，使用摄像头采集了单株和多株的移栽机栽植部件辣椒苗图像，构建辣椒苗数据集；其次在YOLO v5n神经网络架构的基础上，使用Ghost卷积替换普通卷积，并插入改进后的FastGhost模块和SimAMGhost模块，有效降低模型的运算量和计算延迟，提高检测速度；引入EMA注意力机制，提高对重要细节信息的注意程度，改善模型对高度重叠的多株辣椒苗图像识别效果，解决了辣椒苗的部分多检和漏检问题；最终使用Shape-IoU损失函数替换CIoU损失函数，消除边界框自身形状对边界框回归的影响，提高边界框回归准确度。实验结果表明，与YOLO v5n相比，改进后的YOLO v5n-GE模型的检测平均精度均值为95.3%，比原模型提高0.3个百分点，模型参数量和计算量分别缩小52.5%和51.2%，检测速度提升12.2%。与当前YOLO系列主流模型相比，YOLO v5n-GE能够在大幅度减少参数量和运算量的情况下，保持较高的检测精度，证明了改进算法的有效性，可为硬件资源有限的移栽机栽植部件中的辣椒苗识别工作提供技术支持。

Abstract:

Aiming to address key operational issues in automatic transplanting machines, such as missed planting, seedling blockages and abnormal seedling planting state, an optimized lightweight detection model, YOLO v5n-GE, was proposed for real-time monitoring of seedling conditions within transplanting equipment. The research began by collecting images of both single and multiple pepper seedlings under varying lighting conditions (front and backlighting) using a camera. To reduce computational load and latency, traditional convolutions were replaced with Ghost convolutions on the basis of YOLO v5n model, and the main feature extraction modules were substituted with improved FastGhost and SimAMGhost modules. EMA attention mechanism was applied to enhance the network’s focus on important detail information, effectively improving the model’s recognition results for highly overlapping pepper seedlings, reducing sensitivity to occlusions, and increasing recognition accuracy. Additionally, Shape-IoU loss was used to replace CIoU loss, addressing the influence of the bounding box shape on bounding box regression and improving bounding box regression accuracy. Experimental results on the self-built dataset demonstrated that the improved YOLO v5n-GE model achieved an mAP of 95.3%, representing a 0.3 percentage points improvement over the original model. The model’s parameter count and computational load were reduced by 52.5% and 51.2%, respectively, detection speed was increased by 12.2%. These enhancements enabled efficient, real-time detection of pepper seedlings while maintaining high accuracy, demonstrating the improved algorithm’s effectiveness. The research result can provide technical support for seedling recognition in transplanting machine components with limited hardware resources.