基于SGD-YOLO模型的黄瓜霜霉病检测

Cucumber Downy Mildew Detection Based on SGD-YOLO Model

摘要:

为实现复杂环境下黄瓜霜霉病快速定位和精准检测，针对黄瓜霜霉病小样本及小目标问题，基于YOLO v8n提出一种黄瓜霜霉病检测改进模型SGD-YOLO（SimAM Guide-Fusion Dysample-YOLO）。以黄瓜霜霉病叶片为研究对象，通过将显著性检测FT算法引导CutMix方法进行数据增强，结合迁移学习的训练方式，缓解样本数量少带来的过拟合影响。SGD-YOLO在YOLO v8n的基础上引入无参的轻量级模块SimAM（A simple, parameter-free attention module），加强重要特征传播，提高网络整体性能；并采用轻量动态上采样器DySample增强上采样行为，提升病害小目标检测效果；采用CGFM模块（Context guide fusion module）代替Concat模块，通过基于坐标注意力机制（Coordinate attention）实现更精准的多尺度特征融合，优化病害区域的特征提取；损失函数替换为WIoUv3，提供梯度增益分配策略，提高模型泛化性能。结果表明，在增强后的数据集上检测精确率较原数据集提高12.0个百分点，进行迁移学习后检测精确率进一步提高5.3个百分点；改进SGD-YOLO检测精确率为84.6%，平均精度均值（mAP50）达到93.9%，相较于原模型分别提高7.4、9.5个百分点。研究结果对于小样本情况下蔬菜病害检测方法具有较好借鉴作用。

Abstract:

Aiming to achieve rapid localization and precise detection of cucumber downy mildew under complex environmental conditions, an improved detection model named SGD-YOLO (SimAM Guide-Fusion DySample-YOLO) was proposed, based on the YOLO v8n deep learning framework. Addressing the challenges of small sample size and small target detection in cucumber downy mildew, cucumber leaf images infected by the disease were used as research objects. A data augmentation strategy was employed by combining the FT saliency detection algorithm to guide the CutMix method, and a transfer learning approach was adopted to alleviate overfitting caused by limited training data. On top of the YOLO v8n baseline, SGD-YOLO integrated a parameter-free lightweight attention module a simple, parameter-free attention module (SimAM) to enhance the propagation of key features and improve overall network performance. It also employed the dynamic lightweight upsampling module DySample to strengthen upsampling behavior and improve the detection of small diseased targets. In addition, the traditional Concat operation was replaced with the context guide fusion module (CGFM), which leveraged coordinate attention (CoordAtt) to achieve more precise multi-scale feature fusion and better lesion feature extraction. The loss function was replaced with WIoUv3, which incorporated a gradient gain allocation strategy to enhance the model’s generalization ability. Experimental results showed that the augmented dataset improved detection accuracy by 12.0 percentage points over the original dataset, and transfer learning further improved accuracy by 5.3 percentage points. The improved SGD-YOLO achieved an overall detection accuracy of 84.6% and a mean average precision (mAP) of 93.9%, outperforming the baseline model by 7.4 percentage points and 9.5 percentage points, respectively. The research result can provide a valuable reference for small-sample plant disease detection tasks in real-world agricultural applications.