Authors

Abstract

This research aimed to develop a robust crop row detection framework for tobacco fields, where existing methods have not addressed the unique challenges of stratified harvesting, which progressively alters plant morphology from dense lower canopy structures to sparse upper leaf arrangements across multiple collection cycles. CropRowSeg-YOLO (CRS-YOLO), an instance segmentation framework built on YOLOv11, was developed and integrates three core innovations: a Spatial Enhanced Calibration Block (SECB) utilizing depth-wise separable convolutions for spatial feature enhancement; a Hierarchical Adaptive Segmentation Head (HASH) employing asymmetric kernels for long-range dependency capture; and an Enhanced Post-processing Algorithm (EPA) for mask-to-trajectory conversion. Experimental validation in tobacco fields demonstrated 98.4% AP@50 for instance segmentation, with 96.6% precision and 94.2% recall. The model requires 5.1 GFLOPs and 2.7 M parameters, processing individual frames in 2.8 ms, with the complete pipeline executing in 16.99 ms at a resolution of 1920 × 1080.

Abstract in Chinese

本研究旨在开发一种适用于烟草田的稳健作物行检测框架。现有方法无法应对分层收割带来的独特挑战，分层收割会导致植物形态在多个收割周期内逐渐发生变化，从密集的下层冠层结构过渡到稀疏的上层叶片排列。我们开发了 CropRowSeg-YOLO (CRS-YOLO)，这是一个基于 YOLOv11 的实例分割框架，它集成了三项核心创新：利用深度可分离卷积进行空间特征增强的空间增强校准块 (SECB)、采用非对称核进行长程依赖性捕获的分层自适应分割头 (HASH) 以及用于掩码到轨迹转换的增强型后处理算法 (EPA)。在烟草田进行的实验验证表明，实例分割的 AP@50 为 98.4%，精确率为 96.6%，召回率为 94.2%。该模型需要 5.1 GFLOPs 和 2.7M 个参数，处理单个帧需要 2.8 毫秒，整个处理流程在 1920×1080 分辨率下执行时间为 16.99 毫秒