基于改进YOLO v8n的轻量化烟苗移栽质量检测方法

Lightweight Tobacco Seedling Transplanting Quality Detection Method Based on Improved YOLO v8n

摘要:

为提高烟苗移栽机移栽过程烟苗移栽质量识别精度和检测速度，解决设备计算资源有限的问题，本文提出一种改进YOLO v8n的轻量化识别模型，实现烟苗移栽质量检测。在骨干网络（Backbone）中引入基于动态卷积C2f_Ghost(CSP Bottleneck with 2 convolutions_ghost)模块替换C2f(CSP Bottleneck with 2 convolutions)模块，保持精度的同时减小模型内存占用量和参数量，使模型更好地部署在移动设备上。引入ODConv(Omni-dimensional dynamic convolution)替换传统卷积堆叠结构，以新的卷积层替换传统的卷积层，减少因传统卷积层堆叠导致的梯度消失、模型退化，增强了烟苗与地面的判别能力。添加EMA(Efficient multi-scale attention)注意力机制，通过动态特征加权、长程依赖建模与轻量化计算的核心优势，解决烟苗移栽质量检测田间背景干扰强、关键特征尺度差异大、实时检测需求高的问题，提升网络对多尺度目标检测的适应性和准确度。试验结果表明，在自制数据集上改进YOLO v8n模型相比YOLO v8n基线网络模型，精确率、召回率、mAP0.5、mAP0.5-0.95、检测速度分别提高11.3%、17.4%、11.4%、11.8%、4.62%，并且模型内存占用量、参数量和浮点运算量分别下降17.1%、18.3%、31.7%。研究结果为农业作物表型检测模型轻量化设计提供了可行思路，尤其对烟苗移栽质量量化指标的精准提取具有实践价值，推进现代烟草农业数字化管控的落地应用。

Abstract:

Aiming to improve the accuracy and speed of evaluating tobacco seedling transplanting quality while addressing the limited computational resources of agricultural equipment, an improved, lightweight YOLO v8n model was proposed. Firstly, the conventional C2f module in the backbone network was replaced with a C2f_Ghost module. This modification significantly reduced memory footprint and parameter count without compromising accuracy, thereby facilitating deployment on mobile and edge devices. Next, omni-dimensional dynamic convolution (ODConv) was introduced to replace traditional convolutional stacks. This mitigated vanishing gradients and model degradation while enhancing the network's ability to distinguish seedlings from complex soil backgrounds. Furthermore, an efficient multi-scale attention (EMA) mechanism was integrated. Leveraging dynamic feature weighting, long-range dependence modeling, and lightweight computation, EMA effectively overcame challenges such as strong field background interference, significant scale variations of key features, and stringent real-time detection requirements. Experimental results on a custom dataset demonstrated that, compared with the baseline YOLO v8n, the proposed model improved precision, recall, mAP0.5, mAP0.5-0.95, and detection speed by 11.3%, 17.4%, 11.4%, 11.8%, and 4.62%, respectively. Concurrently, memory usage, parameter count, and floating-point operations (FLOPs) were decreased by 17.1%, 18.3%, and 31.7%, respectively. This research can provide a highly viable strategy for designing lightweight models in crop phenotypic detection, particularly for the precise quantitative assessment of transplanting quality, thereby advancing the digital management of modern tobacco agriculture.