基于改进YOLOv8n算法的水稻叶片病害检测

Detection of Rice Leaf Diseases Based on Improved YOLOv8n Algorithm

针对水稻病害检测精度低、易发生漏检和误检等问题,提出了改进YOLO第8常规版(you only look once version 8 normal,YOLOv8n)水稻叶片病害识别检测方法。以原始YOLOv8n算法为基准,在主干网络中使用下采样操作卷积(a convolutional block for down-sampling,ADown)模块,减少特征信息的丢失,并在主干网络与颈部网络之间引入基于挤压-激励(squeeze-and-excitation,SE)的注意力机制模块,提高网络的特征融合能力;同时,设计出共享参数检测头,增加检测任务的感受野;此外,使用顺序证据加权插值算法的交并比(weighted interpolation of sequential evidence for intersection over union,WIoU)损失函数,进一步提升网络的检测性能。在水稻病害数据集上进行大量实验,结果表明,与原始YOLOv8n算法相比,改进YOLOv8n算法的精确率和平均精确率均值分别提升了6.59%和6.86%;改进YOLOv8n算法能够满足水稻叶片病害识别对速度及检测精度的需求,同时提高了对小目标和密集目标的检测能力,从而减少了漏检和误检的情况。改进YOLOv8n算法与目前主流算法相比在检测速度和精度上具有一定优势,检测速度是YOLO第7版(YOLO version 7,YOLOv7)算法的3.61倍,平均精确率均值提高了8.63%。该研究能够为水稻智能化种植管理提供一定的参考。

To address the issues of low detection accuracy,missed detection,and false detection in rice disease identification,an improved method based on you only look once version 8 normal(YOLOv8n)for rice leaf disease recognition and detection was proposed.Original YOLOv8n algorithm was utilized as the baseline model,with a convolutional block for down-sampling(ADown)module employed in the backbone network to reduce the loss of feature information.Additionally,a squeeze-and-excitation(SE)attention mechanism module was introduced between the backbone and neck networks to enhance the network′s feature fusion capability.A shared parameter detection head was also designed to increase the receptive field for detection tasks.Furthermore,the weighted interpolation of sequential evidence for the intersection over union(WIoU)loss function was used to improve the network′s detection performance further.Extensive experiments were conducted on a rice disease dataset.The results indicated that,compared to the original YOLOv8n algorithm,the improved YOLOv8n algorithm achieved a 6.59%increase in precision and a 6.86%increase in mean average precision.The improved YOLOv8n model satisfied the requirements for the speed and accuracy of rice leaf disease identification,while also enhancing the detection capability for small and dense targets,thereby reducing the incidence of missed and false detection.The improved YOLOv8n algorithm demonstrated certain advantages in detection speed and accuracy over current mainstream models,and the detection speed was 3.61 times that of YOLO version 7(YOLOv7)algorithm,achieving an 8.63%increase in mean average precision.This study is expected to have significant reference value in intelligent rice cultivation management.