Abstract: The images from mountainous areas are of many miscellaneous types and are captured remotely by Unmanned Aerial Vehicles, making it difficult to accurately recognize the pine wilt disease. Aiming to solve this problem, this paper proposes a twolevel fusion deep learning model with an attentional mechanism for the classification and recognition of pine wilt disease. Firstly, the VGG16 model is used to classify and “slim” the captured images. Then, the output images containing infected pine wilt disease from the first stage are inputted into the improved YOLOv5 object recognition model. The model further expands the receptive field by introducing the attention mechanism module so that the infected pine wilt disease can be accurately recognized. Finally, the proposed method is compared with other classical deep learning models. The results of comparative experiments show that the proposed twolevel fusion deep learning model based on VGG16 and improved YOLOv5 has the best recognition effect with 85.58% recognition accuracy, which is higher than the other four twolevel fusion deep learning models. The proposed approach can not only improve the accuracy but also solves the problem that manual image classification is needed before pine wilt disease is recognized.

Key words: deep learning, image recognition, pine wilt disease, YOLOv5, VGG16, artificial intelligence(AI)

摘要： 针对无人机遥感获取的山区图像种类多且杂、患松材线虫病松树精准识别难的问题，提出一种带注意力机制的分类和识别两级深度学习模型的松材线虫病识别方法。首先，该方法利用VGG16模型进行图像分类和“瘦身”。然后，将第一级输出的含有患松材线虫病的图像输入到改进的YOLOv5目标识别模型中，该模型通过引入注意力机制模块，进一步扩大感受野，从而对患松材线虫病松树进行精准识别。最后，将所提出的方法和其他经典深度学习模型进行对比试验。结果表明：所提出的基于VGG16和改进的YOLOv5的两级融合深度学习模型的识别效果最好，识别准确率为85.58%，高于其他四种两级融合深度学习模型的识别准确率。所提出的方法不仅提高识别准确率，且解决以往在进行松材线虫病识别前需要人工进行图像分类的问题。

关键词: 深度学习, 图像识别, 松材线虫病, YOLOv5, VGG16, 人工智能(AI)