基于改进YOLO 11s的轻量化大豆炭疽病孢子检测方法

Lightweight Soybean Anthracnose Spore Detection Method Based on Improved YOLO 11s

摘要:

大豆炭疽病由病原体平头炭疽菌侵染引起，其分生孢子可随雨水迅速传播，严重威胁大豆产量与品质，孢子检测是病害早期监测与精准防控的重要基础。针对大豆炭疽病孢子目标尺度小、现有检测模型参数量大、浮点运算量高且难以满足边缘设备部署需求等问题，本文提出一种基于改进YOLO 11s的轻量化大豆炭疽病孢子检测模型HAD-YOLO。主干网络中引入ADown下采样模块，利用多路径池化和通道分离机制，在降低模型复杂度的同时增强孢子边缘与细节特征提取能力;颈部网络采用高级筛选特征金字塔网络（High-level screening feature pyramid network，HS-FPN）替代原有结构PANet，通过通道压缩与动态筛选融合机制强化多尺度特征表达，提升孢子小目标特征表达与定位能力;此外，检测头设计为轻量化共享卷积检测头（Lightweight shared convolutional detection head，LSCD），借助分组卷积与权重共享策略，进一步减少模型参数量与计算量，同时保持低倍率显微场景下孢子定位与分类性能;边界框回归损失函数采用WIOU v3替代CIoU（Complete intersection over union），通过动态聚焦机制减弱低质量样本梯度干扰，提升模型训练稳定性与泛化能力。自建数据集试验结果表明，HAD-YOLO模型精确率、召回率与mAP50分别达到92.3%、85.3%和90.9%，较基线模型分别提升3.4、3.0、4.1个百分点，模型参数量、浮点运算量和模型内存占用量分别降低46.8%、36.6%和46.4%。树莓派4B平台部署测试结果表明，模型单幅图像平均推理时间为57.5ms，具备较好的实时检测能力。本文所提方法在保证较高检测精度的同时实现了模型轻量化，为大豆炭疽病孢子显微图像智能检测和边缘端部署提供方法基础。

Abstract:

Soybean anthracnose is caused by the pathogen Colletotrichum truncatum, and its conidia, the soybean anthracnose spores can spread rapidly with rain, seriously threatening soybean yield and quality.Spore detection is a critical foundation for early disease monitoring and precise control.To address issues such as the small target scale of soybean anthracnose spores, large number of parameters in existing detection models, high computational complexity, and difficulty of meeting edge device deployment requirements, a lightweight soybean anthracnose spore detection model, HAD-YOLO, was proposed based on the improved YOLO 11s.Firstly, the backbone network introduced the ADown downsampling module, using multi-path pooling and channel separation mechanisms to enhance edge and detail feature extraction of spores while reducing model complexity.Secondly, the neck network adopted the high-level screening feature pyramid network (HS-FPN) instead of the original PANet structure, strengthening multi-scale feature representation through channel compression and dynamic screening fusion mechanisms and improving the encoding and localisation capability for small spore targets.Additionally, the detection head was designed as a lightweight shared convolutional detection head (LSCD), further reducing model parameters and computational load through grouped convolution and weight-sharing strategies while maintaining spore localisation and classification performance under low-magnification microscopic scenarios.Finally, the bounding box regression loss function used WIOU v3 instead of complete intersection over union (CIoU), mitigating low-quality sample gradient interference through a dynamic focusing mechanism, enhancing model training stability and generalisation ability.Experimental results on a self-built dataset showed that the HAD-YOLO model achieved precision, recall, and mAP50 of 92.3%, 85.3%, and 90.9%, respectively, improving over the baseline model by 3.4, 3.0, and 4.1 percentage points, while reducing model parameters, computational complexity, and weight file size by 46.8%, 36.6%, and 46.4%, respectively.Deployment tests on the Raspberry Pi 4B platform showed an average inference time of 57.5ms per image, demonstrating favourable real-time detection capability.The research result indicated that this method achieved model lightweighting while ensuring detection precision, providing a methodological foundation for intelligent microscopic image detection and edge deployment of soybean anthracnose spores.