DA-YOLO：基于双重注意力的松枯病检测模型

DA-YOLO: A Detection Model for Pine Wilt Disease Based on Dual Attention

摘要: 松枯病(PWD)是一种传播迅速、杀伤力极强的森林病害，对我国森林生态安全构成严重威胁，并造成巨大的林业经济损失。考虑到我国森林面积广阔，人工巡查监测难度大且成本高，因此利用无人机遥感技术监测病树成为控制松枯病传播的有效途径。尽管目前松枯病检测算法取得了相对较好的性能，但由于松枯病的强传染性，检测效果仍需进一步提高。基于此，本文提出了一种基于YOLOv5的双重注意力混合模型——DA-YOLO，用于更有效地检测病害树木区域。该算法使用基于自注意力的CoT模块加强骨干特征网络的提取能力，并结合ECA注意力机制整体提升定位精度。实验结果显示，在使用PWD遥感数据集时，该模型的AP@0.5:0.95较之基线提高了5.2个百分点。并将本文提出的算法DA-YOLO与Faster R-CNN、RetinaNet、YOLOv5、YOLOv6、YOLOx、YOLOv7等算法的模型复杂度和精度进行对比，并分析了它们在检测松枯萎线虫树木方面的效果。实验结果表明，本文提出的DA-YOLO模型在检测方面具有明显的优势。

Abstract: This Pine wilt disease (PWD) is a rapidly spreading and highly lethal forest disease that poses a serious threat to China’s forest ecological security and causes significant economic losses in forestry. Considering the vast forest area in China, the difficulty and high cost of manual inspection and monitoring, the use of drone remote sensing technology to monitor diseased trees has become an effective way to control the spread of pine wilt disease. Although the current pine wilt disease detection algorithm has achieved relatively good performance, the detection effect still needs to be further improved due to the strong infectivity of pine wilt disease. Based on this, this article proposes a dual attention hybrid model called DA-YOLO based on YOLOv5, which is used to more effectively detect areas of diseased trees. This algorithm uses a self attention based CoT module to enhance the extraction ability of the backbone feature network, and combines ECA attention mechanism to improve the overall positioning accuracy. The experimental results show that when using the PWD remote sensing dataset, the model’s AP@0.5 0.95 is 5.2 percentage points higher than the baseline. And the model complexity and accuracy of the algorithm DA-YOLO proposed in this article were compared with Faster R-CNN, RetinaNet, YOLOv5, YOLOv6, YOLOx, YOLOv7 and other algorithms, and their performance in detecting pine wilt nematode trees was analyzed. The experimental results show that the DA-YOLO model proposed in this paper has significant advantages in detection.