基于YOLO v8-PicoDet的水稻叶片病害识别

Rice Leaf Diseases Identification Based on YOLO v8-PicoDet

摘要:

针对水稻叶片病害在田间环境下多样复杂、小目标检测困难及智能设备部署困难等难题,本文提出一种融合YOLO v8n模型与PicoDet模块的轻量化目标检测模型YOLO v8-PicoDet,用于实现水稻叶片病害的高效检测。模型以ESNet（Efficient symmetric network）作为主干网络,引入SimAM（Simple attention module）注意力机制有效降低了模型参数量和计算复杂度,参数量降低1.6×10^6,浮点运算量下降2.1×10^9。通过构建CSP-PAN（Cross stage partial path aggregation network）特征融合结构,提升了对小目标病斑的检测能力。在检测头中引入CBAM（Convolutional block attention module）注意力机制,加强了模型对复杂病斑特征的表征能力。最后,通过组合优化GIoU Loss与Distribution Focal Loss损失函数,从分类、回归和置信度多个维度优化模型检测与分类性能。改进模型部署在Raspberry Pi 4B上运行,运行帧率为28.5 f/s,克服了原始YOLO v8n模型难以在便携式设备上运行的不足。mAP@0.5和mAP@0.5:0.95达93.8%、76.9%,较YOLO v8n模型提升0.9、1.3个百分点,模型性能优于现有主流检测模型。消融试验结果进一步表明,本文所设计各优化模块对模型性能提升均有正向贡献,使模型在计算资源有限的便携式设备上获得较同类模型更高的检测精度和实时性,为智慧农业领域水稻病害的实时智能检测提供了新思路和技术支持。

Abstract:

The detection and identification of rice leaf diseases hold significant importance in agricultural production. To address the challenges of diverse and complex field environments, difficulties in detecting small disease targets, and deployment constraints of intelligent devices for rice leaf disease detection, a lightweight object detection model named YOLO v8-PicoDet was proposed, which integrated the advantages of YOLO v8n and PicoDet, aiming to achieve efficient detection of rice leaf diseases. Firstly, the model took efficient symmetric network (ESNet) as its backbone and introduced the simple attention module (SimAM) to reduce the number of model parameters and the computational cost. As a result, the number of parameters was reduced by 1.6×10^6 and the FLOPs was decreased by 2.1×10^9. Secondly, the cross stage partial path aggregation network (CSP-PAN) feature fusion was constructed, which enhanced the detection ability for small target lesions. Again, the convolutional block attention module (CBAM) was introduced into the detection head, enhancing the model's ability to represent the characteristics of complex lesions. Finally, the GIoU Loss and Distribution Focal Loss functions were combined and optimized to optimize the model detection and classification performance from multiple dimensions such as classification, regression and confidence. The experiment was conducted based on the Raspberry Pi 4B platform, and the results showed that the optimized model can achieve an inference speed of 28.5 f/s, overcoming the deficiency that many models were difficult to run on portable devices. Meanwhile, the model achieved 93.8% mAP@0.5 and 76.9% mAP@0.5:0.95, which was 0.9 percentage points and 1.3 percentage points higher than that of the original YOLO v8n model, and its performance was superior to existing mainstream detection models.