基于空间深度转换域自适应学习的小麦赤霉病无监督检测方法

Unsupervised Detection of Wheat Scab Based on Space-to-depth Domain Adaptation Learning

摘要:

利用无人机遥感监测小麦赤霉病时，随着无人机飞行高度的增加，所采集到的图像中病斑尺寸也会越来越小。这些小目标会增加数据标注的难度，且需要耗费更多的时间和人工成本。针对该问题，本文提出了一种无监督的无人机小麦图像赤霉病检测方法 - 空间深度转换域自适应检测网络 (Space-to-depth domain adaptation detection network,SPD-DANet), 通过对抗思想学习源域数据的知识，并将其迁移至无标注的目标域从而实现小麦赤霉病无监督检测。首先，针对无人机小麦图像中赤霉病病斑较小的问题，设计了空间深度转换特征提取器 (Space-to-depth feature extractor,SPD-FE), 将无人机图像的空间信息转换为深度信息，使得网络能更加有效地学习小目标病斑的特征。其次，利用 SPD-FE 构建基于对抗的分类自适应和边界框回归自适应模块，让检测网络分别从分类和边界框回归 2 个步骤依次进行域自适应学习，从而利用源域数据的知识实现目标域的无监督检测。实验结果表明，本文提出的方法相他目标测方法，DETR、YOLOv5 等，检测精度 (AP50) 提高 5.3~20.5 个百分点，在小麦赤霉病的无监督检测任务中表现最佳，且与基线网络相比其检测精度 (AP50) 提高 6.5 个百分点。本文的研究结果能够为小麦赤霉病的无监督检测提供技术支撑。

Abstract:

When using UAV remote sensing to monitor wheat scab, the size of the object in the captured images decreases as the UAV's flight altitude increases. This makes data annotation more challenging, requiring more time and labor. To address this issue, an unsupervised UAV-based wheat scab detection method was proposed. The method proposed a space-to-depth domain adaptation detection network (SPD - DANet), which learned the knowledge of the source domain data through the adversarial idea and transfered it to the unlabeled target domain thereby realizing unsupervised wheat scab detection. Firstly, to tackle the issue of small scab lesions in UAV -captured wheat images, a spatial-to-depth feature extractor (SPD - FE) that transformed spatial information into depth information was designed. This enabled the network to learn features of small scab objects more effectively. Secondly, using SPD - FE, adversarial-based classification adaptation and bounding box regression adaptation modules for the detection network were constructed. This enabled domain adaptation learning sequentially in classification and bounding box regression steps, leveraging source domain knowledge for unsupervised detection in the target domain. The experimental results showed that the proposed method improved the detection accuracy ( AP50) by 5.3 ~ 20. 5 percentage points compared with other object detection methods such as DETR, YOLO v5, etc., and performed the best on the unsuperrised detection task of wheat scab, and its detection accuracy AP50 was improved by 6.5 percentage points compared with that of the baseline network. The research can provide some support and help for the unsupervised detection of wheat scab.