基于改进YOLO v4的荔枝病虫害检测模型

为实时准确地检测到自然环境下背景复杂的荔枝病虫害,本研究构建荔枝病虫害图像数据集并提出荔枝病虫害检测模型以提供诊断防治。以YOLO v4为基础,使用更轻、更快的轻量化网络GhostNet作为主干网络提取特征,并结合GhostNet中的核心设计引入更低成本的卷积Ghost Module代替颈部结构中的传统卷积,得到轻量化后的YOLO v4-G模型。在此基础上使用新特征融合方法和注意力机制CBAM对YOLO v4-G进行改进,在不失检测速度和模型轻量化程度的情况下提高检测精度,提出YOLO v4-GCF荔枝病虫害检测模型。构建的数据集包含荔枝病虫害图像3725幅,其中病害种类包括煤烟病、炭疽病和藻斑病3种,虫害种类包括毛毡病和叶瘿蚊2种。试验结果表明,基于YOLO v4-GCF的荔枝病虫害检测模型,对于5种病虫害目标在训练集、验证集和测试集上的平均精度分别为95.31%、90.42%和89.76%,单幅图像检测用时0.1671s,模型内存占用量为39.574MB,相比改进前的YOLO v4模型缩小84%,检测速度提升38%,在测试集中检测平均精度提升4.13个百分点,同时平均精度比常用模型YOLO v4-tiny、EfficientDet-d2和Faster R-CNN分别高17.67、12.78、25.94个百分点。所提出的YOLO v4-GCF荔枝病虫害检测模型能够有效抑制复杂背景的干扰,准确且快速检测图像中荔枝病虫害目标,可为自然环境下复杂、非结构背景的农作物病虫害实时检测研究提供参考。

Detection of Litchi Diseases and Insect Pests Based on Improved YOLO v4 Model

In order to accurately detect litchi diseases and insect pests with complex background in natural environment in real time, the data set of litchi diseases and insect pests was constructed and the detection model of litchi diseases and insect pests was proposed for diagnosis and control. Based on YOLO v4, GhostNet, the lighter and faster lightweight network, was used as the backbone network to extract features. According to the core design of GhostNet, Ghost Module, a lower cost convolution, was used to replace the traditional convolution in the neck structure. Based on the lightweight YOLO v4-G model, the feature fusion method and attention mechanism called CBAM were used to improve the YOLO v4-G. The detection accuracy was improved without losing the detection speed and the lightweight degree of the model. Finally, the YOLO v4-GCF detection model of litchi diseases and insect pests was proposed. The dataset contained 3725 images of litchi diseases and insect pests. Litchi diseases included sooty mold, anthracnose and algal spot. Litchi insect pests included leaf mite and Dasineura sp. The experimental results showed that the average accuracy of five kinds of diseases and insect pests targets detected by YOLO v4-GCF detection model in train set, validation set and test set was 95.31%, 90.42% and 89.76%, respectively. The detection time of a single image was 0.1671s, and the size of the model was 39.574MB. Compared with the YOLO v4, the model size was reduced by 84%, the detection speed was increased by 38% and the average accuracy in the test set was improved by 4.13 percentage points. At the same time, the average accuracy was 17.67,12.78 and 25.94 percentage points higher than those of YOLO v4-tiny, EfficientDet-d2 and Faster R-CNN, respectively. The proposed YOLO v4-GCF detection model of litchi diseases and insect pests can effectively inhibit the interference of complex background, and accurately and quickly detect targets of litchi diseases and insect pests in the images, which can provide reference for crop diseases and insect pests detection research with complex and unstructured background in natural environment.