基于无人机RGB图像与改进YOLO v5s的宿根蔗缺苗定位方法

Method for Locating Missing Ratoon Sugarcane Seedlings Based on RGB Images from Unmanned Aerial Vehicles and Improve YOLO v5s

摘要:

针对预切种式双芽蔗段横向补种机缺少整体的缺苗数据，导致补种效率不高等问题，提出了一种基于无人机RGB图像的宿根蔗缺苗定位方法。首先，通过无人机快速采集实际田间宿根蔗幼苗的高分辨率图像，将航拍大图（分辨率为5472像素×3648像素）切分成多幅子图并进行数据增强，从而构建宿根蔗幼苗数据集；其次，在YOLO v5s的基础上引入P2小目标特征层和DyHead模块，提高对幼苗小目标的检测准确性，并在训练过程引入图像加权策略解决样本数量不平衡问题，进一步提高被遮挡幼苗的检测精度；然后，在切片辅助推理框架中引入改进模型训练权重，在大尺寸田间图像中实现宿根蔗幼苗的检测；最后，构建以改进的DBSCAN聚类算法和PCA拟合算法为核心的作物行识别算法，在作物行线上定位缺苗位置。试验结果表明，改进宿根蔗幼苗检测模型在子图上的平均检测精度为96.8%，在大图上的识别精确率和召回率为94.5%和91.8%，检测时间为0.32s。基于检测的位置坐标信息利用作物行识别算法实现分垄，作物行聚类准确率达到100%，拟合的作物行中心线角度平均误差为0.2455°，作物行中心线上缺苗位置识别的精确率和召回率为91.9%和97.1%，平均定位误差为9.73像素。该方法可用于大尺寸复杂田间图像上的宿根蔗智能缺苗定位，为补种作业提供技术支持，对延长宿根年限、提高甘蔗产量具有重要意义。

Abstract:

In response to the lack of specific missing seedling data for the transverse replanting machine of pre-cut double bud sugarcane segments, resulting in poor replanting efficiency, a method for locating missing ratoon sugarcane seedlings based on UAV RGB images was proposed. Firstly, high-resolution images of ratoon sugarcane seedlings in the field were rapidly captured by using UAVs, which were then segmented into multiple sub-images and subjected to data augmentation to construct a dataset. Secondly, enhancements to the YOLO v5s model involved the introduction of P2 small target feature layers and DyHead modules to improve the detection accuracy of small seedling targets. Additionally, an image weighting strategy was employed during training to address sample imbalance issues and further improve detection accuracy, especially for occluded seedlings. Subsequently, a framework incorporating sliced-assisted inference facilitated the detection of ratoon sugarcane seedlings in large-scale field images by using the trained model. Finally, a row recognition algorithm based on an improved DBSCAN clustering algorithm and PCA fitting algorithm was developed to locate missing seedling positions along crop rows. Experimental results demonstrated that the improved ratoon sugarcane seedling detection model achieved an average detection accuracy of 96.8% on sub-images and recognition precision and recall rates of 94.5% and 91.8%, respectively, on large-scale images, with a detection time of 0.32s. Utilizing the detection coordinates, the row recognition algorithm achieved 100% clustering accuracy, with an average angular error of 0.2455° for fitted row angles, and precision and recall rates of 91.9% and 97.1%, respectively, for missing seedling detection along rows. This method can be applied to intelligent missing seedling localization in large-scale, complex field images of ratoon sugarcane, providing technical support for replanting operations and holding significant implications for extending ratoon lifespan and increasing sugarcane yield.