基于Faster-NAM-YOLO的黄瓜霜霉病菌孢子检测

Quantitative Detection of Cucumber Downy Mildew Spores at Multi-scale Based on Faster-NAM-YOLO

黄瓜霜霉病由古巴假霜霉病菌孢子通过侵染引起,严重影响了黄瓜的品质和产量;病菌孢子数量与病情严重度相关,因此建立快速、简便和高效的病菌孢子定量检测方法,实现黄瓜霜霉病防治关口前移。基于YOLO v5模型提出了一种基于Faster-NAM-YOLO的黄瓜霜霉病菌孢子定量检测模型,该模型首先提出了一种特征提取模块C3_Faster,使用C3_Faster替换YOLO v5中的C3模块,有效降低了模型参数计算量和模型深度,提升了对黄瓜霜霉病菌孢子检测速度和精度;其次在主干网络中加入了NAM注意力模块,通过应用权重稀疏性惩罚抑制不显著权重,进而提高模型的特征提取能力和计算效率;最后实现了对黄瓜霜霉病菌孢子的定量检测。实验结果表明,Faster-NAM-YOLO模型在测试集上mAP@0.5和mAP@0.5:0.95分别达到95.80%和60.90%,对比原始YOLO v5模型分别提升1.80、1.20个百分点,较原始YOLO v5模型内存占用量和每秒浮点运算次数分别减少5.27 MB和1.49×10^(10);通过与YOLO v3、THP-YOLO v5、YOLO v7、YOLO v8、Faster RCNN、SSD目标检测模型对比,Faster-NAM-YOLO在检测精度、模型内存占用量、每秒浮点运算次数和推理时间方面均具有显著优势;在1 200像素×1 200像素、1 500像素×1 500像素和1 800像素×1 800像素3种不同分辨率尺度及不同图像数量下进一步验证了Faster-NAM-YOLO模型具有较强的鲁棒性和泛化能力。

Cucumber downy mildew is caused by the spores of cucumber downy mildew from Cuba through infection,which seriously affects the quality and yield of cucumber.The number of the spores is closely related to the severity of the disease.Accordingly,there urgently needs to establish a rapid,simple and efficient quantitative detection method for the spores of cucumber downy mildew,in order to explore the way forward to achieve the control of cucumber downy mildew.Based on YOLO v5 model,an exploratory model was proposed for quantitative detection of cucumber downy mildew spores by Faster-NAM-YOLO.Firstly,a feature extraction module C3_Faster was proposed,which was used to replace the C3 module in YOLO v5,which effectively reduced the calculation amount of model parameters and the depth of the model,and also improved the detection speed and accuracy of cucumber downy mildew spores.Secondly,the NAMAttention module was added to the backbone network and also improved the model's feature extraction ability and computational efficiency by applying weight sparsity penalty to suppress insignificant weights.In the end,the quantitative detection of the spores caused by cucumber downy mildew was realized.Faster-NAM-YOLO model on the test set mAP@0.5 and mAP@0.5:0.95 reached 95.80%and 60.90%,respectively,to compare with the original Y0L0 model.It can be seen that the final results were increased by 1.80 percentage points and 1.20 percentage points,respectively,reducing the model size and FLOPs of the original YOLO v5 model by 5.27M and 1.49×10^(10),respectively.It was found that Faster-NAM-YOLO had significant advantages in detection accuracy,model size,FLOPs,and inference time compared with single stage target detection models such as YOLO v3,THP-YOLO v5,YOLO v7,YOLO v8,Faster RCNN,and SSD.In addition,under the comparison for the three different resolution scales of 1200 pixels×1200 pixels,1500 pixels×1500 pixels and 1800 pixels × 1800 pixels,as well as the different specifications and the varied amount of images,which suggested that the Faster-NAM-YOLO model was further validated to have strong robustness and generalization ability.The research result not only provided a more accurate basis for early online monitoring of cucumber downy mildew,but also laid a foundation for further exploring the relationship between the dynamic changes of spore number and morphological characteristics and the severity of the disease.