Li Xudong, Zhou Yuhong, Liu Jingyan, Wang Linbai, Zhang Jun, Fan Xiaofei
State Key Laboratory of North China Crop Improvement and Regulation, Baoding, China.
College of Mechanical and Electrical Engineering, Hebei Agricultural University, Baoding, China.
Front Plant Sci. 2022 Jul 5;13:899754. doi: 10.3389/fpls.2022.899754. eCollection 2022.
Potato early blight and late blight are devastating diseases that affect potato planting and production. Thus, precise diagnosis of the diseases is critical in treatment application and management of potato farm. However, traditional computer vision technology and pattern recognition methods have certain limitations in the detection of crop diseases. In recent years, the development of deep learning technology and convolutional neural networks has provided new solutions for the rapid and accurate detection of crop diseases. In this study, an integrated framework that combines instance segmentation model, classification model, and semantic segmentation model was devised to realize the segmentation and detection of potato foliage diseases in complex backgrounds. In the first stage, Mask R-CNN was adopted to segment potato leaves in complex backgrounds. In the second stage, VGG16, ResNet50, and InceptionV3 classification models were employed to classify potato leaves. In the third stage, UNet, PSPNet, and DeepLabV3+ semantic segmentation models were applied to divide potato leaves. Finally, the three-stage models were combined to segment and detect the potato leaf diseases. According to the experimental results, the average precision (AP) obtained by the Mask R-CNN network in the first stage was 81.87%, and the precision was 97.13%. At the same time, the accuracy of the classification model in the second stage was 95.33%. The mean intersection over union (MIoU) of the semantic segmentation model in the third stage was 89.91%, and the mean pixel accuracy (MPA) was 94.24%. In short, it not only provides a new model framework for the identification and detection of potato foliage diseases in natural environment, but also lays a theoretical basis for potato disease assessment and classification.
马铃薯早疫病和晚疫病是影响马铃薯种植和生产的毁灭性病害。因此，对这些病害进行精确诊断对于马铃薯农场的治疗应用和管理至关重要。然而，传统的计算机视觉技术和模式识别方法在作物病害检测方面存在一定局限性。近年来，深度学习技术和卷积神经网络的发展为作物病害的快速准确检测提供了新的解决方案。在本研究中，设计了一个结合实例分割模型、分类模型和语义分割模型的集成框架，以实现复杂背景下马铃薯叶片病害的分割和检测。在第一阶段，采用Mask R-CNN在复杂背景下分割马铃薯叶片。在第二阶段，使用VGG16、ResNet50和InceptionV3分类模型对马铃薯叶片进行分类。在第三阶段，应用UNet、PSPNet和DeepLabV3+语义分割模型对马铃薯叶片进行划分。最后，将三个阶段的模型结合起来对马铃薯叶片病害进行分割和检测。根据实验结果，第一阶段Mask R-CNN网络获得的平均精度（AP）为81.87%，精确率为97.13%。同时，第二阶段分类模型的准确率为95.33%。第三阶段语义分割模型的平均交并比（MIoU）为89.91%，平均像素准确率（MPA）为94.24%。简而言之，它不仅为自然环境中马铃薯叶片病害的识别和检测提供了一个新的模型框架，也为马铃薯病害评估和分类奠定了理论基础。
