基于改进YOLO v12与主方向压缩聚类的水稻苗带线提取

Seedling Band Line Extraction Method Based on Improved YOLO v12 and Clustering with Principal Direction Compression

摘要:

水稻秧苗行精准识别对提升水田除草机械的自主导航与对行作业精度至关重要。针对水田中普遍存在的浮萍覆盖、杂草干扰、秧苗缺失及光照不均等问题,本文提出了一种融合目标检测与主方向压缩聚类的秧苗行识别算法。首先构建了集成多尺度门控注意力和轻量化共享检测头的水稻秧苗检测模型（RSD）,有效增强了对秧苗上下文特征提取能力及对复杂背景的抑制能力。进而,构建基于方向性压缩迭代聚类策略,通过自适应地弥合同行秧苗点间纵向间隙,解决传统方法在缺苗场景下苗行断裂问题,实现了从离散检测点到连续苗带线稳健拟合。试验结果表明,秧苗检测中,RSD模型AP0.5达92.5%,检测速度为104.2 f/s,性能优于主流对比模型;在苗带线提取中,该算法在不同缺苗情况下行断裂率与误合并率均低于1%,平均识别角度偏差仅为0.203°,试验结果充分验证了算法性能优势。

Abstract:

Accurate recognition of rice seedling rows is crucial for improving the autonomous navigation and row-following precision of weeding machinery in paddy fields. To solve the problems in paddy fields, such as duckweed coverage, weed interference, missing seedlings, and uneven illumination, a seedling row recognition algorithm that integrated object detection with principal direction compression clustering was proposed. The algorithm firstly constructed a rice seedling detection model, which combined a multi-scale gated attention mechanism with a lightweight shared detection head, effectively enhancing the extraction of seedling contextual features and the suppression of complex backgrounds. Furthermore, an iterative clustering strategy based on directional compression was developed, which adaptively bridged the longitudinal gaps between seedling points within the same row, thereby solving the problem of row breakage faced by traditional methods in seedling gap scenarios and achieving robust fitting from discrete detection points to continuous seedling band lines. Experimental results demonstrated that the proposed RSD model achieved an AP0.5 of 92.5% and a detection speed of 104.2 f/s in the seedling detection task, outperforming mainstream comparison models. The seedling band line extraction method achieved an average angular deviation of 0.203°, with both the row breakage and mis-merging rates below 1% across various seedling gap conditions. Comparative experiments fully validated the performance advantages of this method.