基于改进ATSS模型的水稻叶片病害检测

Rice Leaf Diseases Detection Based on Improved ATSS Model

针对传统水稻病害诊断方法依赖人工、容易误判等缺点,提出一种基于ATSS的水稻叶片病害检测模型。首先收集白叶枯病、胡麻斑病、叶瘟病这三种病害图像,构建水稻叶片病害图像数据集。然后在原ATSS模型的基础上,网络Neck部分采用FPN-CARAFE模块代替特征金字塔网络FPN,以减少上采样过程中的信息损失。同时,为提升模型的检测效果,回归分支的损失函数采用CIoU损失函数代替GIoU。改进ATSS模型的平均精度均值可达74.0%,相比于原ATSS模型提升了3.5%。与模型Retinanet、Faster R-CNN、Cascade R-CNN、FCOS、TOOD相比,改进ATSS模型取得了最高的检测精度,且在检测精度和速度上取得了最高的权衡。实验结果表明,改进后的模型能对水稻叶片病害有效检测。

A rice leaf disease detection model based on ATSS is proposed to address the shortcomings of traditional rice disease diagnosis methods that rely on manual labor and are prone to misjudgment.Firstly,the images of white leaf blight,brown spot and leaf blast were collected to construct rice leaf disease image dataset,then,based on the original ATSS model,the Neck part of the network uses an FPN-CARAFE module instead of FPN module to reduce information loss during the upsampling process;FPN-CARAFE module is used to replace the characteristic pyramid network FPN in the network Neck to reduce the information loss in the up-sampling process based on the original ATSS model.Meanwhile,to improve the detection effect of the model,the loss function of the regression branch adopts the CIoU instead of GIoU.The mean average precision of the improved ATSS model can reach 74.0%,which is 3.5%higher than that of the original ATSS model.Compared with models Retinanet,Faster R-CNN,Cascade R-CNN,FCOS and TOOD,the improved ATSS model has the highest detection accuracy and the highest weight in detection accuracy and speed.The experimental results indicate the improved model can effectively detect rice leaf diseases.