Tomato gray leaf spot is a worldwide disease, especially in warm and humid areas. The continuous expansion of greenhouse tomato cultivation area and the frequent introduction of foreign varieties in recent years have increased the severity of the epidemic hazards of this disease in some tomato planting bases annually. This disease is a newly developed one. Thus, farmers generally lack prevention and control experience and measures in production; the disease is often misdiagnosed or not prevented and controlled timely; this condition results in tomato production reduction or crop failure, which causes severe economic losses to farmers. Therefore, tomato gray leaf spot disease should be identified in the early stage, which will be important in avoiding or reducing the economic loss caused by the disease. The advent of the era of big data has facilitated the use of machine learning method in disease identification. Therefore, deep learning method is proposed to realise the early recognition of tomato gray leaf spot. Tomato growers need to develop the app of image detection mobile terminal of tomato gray leaf spot disease to realise real-time detection of this disease. This study proposes an early recognition method of tomato leaf spot based on MobileNetv2-YOLOv3 model to achieve a good balance between the accuracy and real-time detection of tomato gray leaf spot. This method improves the accuracy of the regression box of tomato gray leaf spot recognition by introducing the GIoU bounding box regression loss function. A MobileNetv2-YOLOv3 lightweight network model, which uses MobileNetv2 as the backbone network of the model, is proposed to facilitate the migration to the mobile terminal. The pre-training method combining mixup training and transfer learning is used to improve the generalisation ability of the model. The images captured under four different conditions are statistically analysed. The recognition effect of the models is evaluated by the F1 score and the AP value, and the experiment is compared with Faster-RCNN and SSD models. Experimental results show that the recognition effect of the proposed model is significantly improved. In the test dataset of images captured under the background of sufficient light without leaf shelter, the F1 score and AP value are 94.13% and 92.53%, and the average IOU value is 89.92%. In all the test sets, the F1 score and AP value are 93.24% and 91.32%, and the average IOU value is 86.98%. The object detection speed can reach 246 frames/s on GPU, the extrapolation speed for a single 416 × 416 picture is 16.9 ms, the detection speed on CPU can reach 22 frames/s, the extrapolation speed is 80.9 ms and the memory occupied by the model is 28 MB. The proposed recognition method has the advantages of low memory consumption, high recognition accuracy and fast recognition speed. This method is a new solution for the early prediction of tomato leaf spot and a new idea for the intelligent diagnosis of tomato leaf spot.

中文翻译：

---

基于MobileNetv2-YOLOv3模型的番茄灰斑病早期识别

番茄灰斑病是一种世界性病害，特别是在温暖潮湿的地区。近年来温室番茄种植面积的不断扩大和国外品种的频繁引进，使得该病害在一些番茄种植基地的流行危害程度逐年增加。本病为新发病。因此，农民在生产中普遍缺乏防控经验和措施；该病常被误诊或未及时防治；这种情况导致番茄减产或歉收，给农民造成严重的经济损失。因此，应及早发现番茄灰斑病，这对于避免或减少该病造成的经济损失具有重要意义。大数据时代的到来促进了机器学习方法在疾病识别中的应用。因此，提出深度学习方法来实现番茄灰斑病的早期识别。番茄种植户需要开发番茄灰斑病图像检测移动端APP，实现对该病害的实时检测。本研究提出了一种基于MobileNetv2-YOLOv3模型的番茄叶斑病早期识别方法，在番茄灰斑病的准确率和实时检测之间取得了很好的平衡。该方法通过引入GIoU边界框回归损失函数，提高了番茄灰叶斑识别回归框的准确率。一个MobileNetv2-YOLOv3轻量级网络模型，使用MobileNetv2作为模型的主干网络，建议方便迁移到移动终端。采用混合训练和迁移学习相结合的预训练方法来提高模型的泛化能力。对四种不同条件下拍摄的图像进行统计分析。通过F1分数和AP值来评价模型的识别效果，并与Faster-RCNN和SSD模型进行实验对比。实验结果表明，该模型的识别效果显着提高。在光照充足且无遮蔽的背景下拍摄的图像测试数据集中，F1分数和AP值分别为94.13%和92.53%，平均IOU值为89.92%。在所有测试集中，F1分数和AP值分别为93.24%和91.32%，平均IOU值为86.98%。GPU上物体检测速度可达246帧/秒，单张416×416图片外推速度为16.9毫秒，CPU上检测速度可达22帧/秒，外推速度为80.9毫秒，内存占用模型为 28 MB。所提出的识别方法具有内存消耗低、识别准确率高、识别速度快等优点。该方法为番茄叶斑病的早期预测提供了新的解决方案，为番茄叶斑病的智能诊断提供了新思路。识别准确率高，识别速度快。该方法为番茄叶斑病的早期预测提供了新的解决方案，为番茄叶斑病的智能诊断提供了新思路。识别准确率高，识别速度快。该方法为番茄叶斑病的早期预测提供了新的解决方案，为番茄叶斑病的智能诊断提供了新思路。