基于改进YOLO v8n的玉米病害检测系统

Maize Disease Detection System Based on Improved YOLO v8n

摘要:

针对YOLO v8n目标检测模型在光照变化显著和背景杂乱条件下识别玉米病害时存在高漏检率和鲁棒性差的问题，本文提出基于改进YOLO v8n的新型玉米病害识别模型DSEMA-YOLO。在YOLO v8n基础上，通过在C2f模块瓶颈前集成EMA全局注意力机制，构建了新型C2f-attention模块，有效减少复杂背景对病害图像识别的干扰，同时完整捕捉上下文信息。通过添加DSConv模块对颈部网络结构进行重构，采用自适应卷积核形状调整机制，能更精准捕捉图像中玉米病害目标特征，缓解光照变化和背景杂乱的影响。最后，通过引入尺度160像素×160像素的特征融合层优化颈部网络结构，提升小目标边界框检测精度，增强浅层与深层信息融合，从而提高识别准确率。头部网络中新增微目标检测模块，并与尺度160像素×160像素特征融合层相连。田间环境采集的玉米病害数据集试验结果表明，改进模型mAP@0.5为92.2%，较YOLO v8n模型提升7.7个百分点。在保持检测速度和较低计算开销的同时有效提升了玉米病害识别精度。

Abstract:

Aiming to address the high missed detection rate and poor robustness of the YOLO v8n target detection model when identifying maize diseases under significant lighting changes and background disorder, a novel maize disease identification model was proposed based on improved YOLO v8n, DSEMA-YOLO.Based on YOLO v8n, by integrating the EMA global attention mechanism before the C2f module bottleneck, a C2f-attention module was constructed, effectively reducing interference from complex backgrounds on defective image recognition while fully capturing contextual information.By adding the DSConv module to reconstruct the neck network structure and adopting an adaptive convolutional kernel shape adjustment mechanism, it can more accurately capture target features of corn diseases in the image, mitigating the effects of lighting changes and background clutter.Finally, by introducing a 160 pixel×160 pixel scale feature fusion layer to optimize the neck network structure, the detection accuracy of small object bounding boxes was improved, and the fusion of shallow and deep information was strengthened, thereby improving recognition accuracy.A micro-object detection module was added to the head network, connected to the 160 pixel×160 pixel scale feature fusion layer.In the experiment, the improved model mAP@0.5 of the maize disease dataset collected in the field environment was 92.2%, an improvement of 7.7 percentage points over that of the YOLO v8n model.While maintaining detection speed and low computational overhead, it effectively improved the accuracy of maize disease identification.