摘要：

无人机影像中的建筑物提取主要面临两大挑战, 首先是容易受到树木和阴影的遮挡, 导致分割错误, 其次是现有方法常忽略建筑物的形态和多分辨率信息。为了应对这些挑战, 在对抗网络的框架内引入双通道并行的生成器策略, 其中一个通道基于形态驱动的小波变换, 专注于捕获建筑物的形态属性, 包括建筑物的轮廓和结构特征, 另一个通道基于DeepLabv3+, 用于处理建筑物的复杂纹理, 包括表面纹理和细节信息, 这种设计使网络可以从多个方面理解影像中的建筑物特征。同时, 为了应对遮挡问题, 提出一种遮挡感知预处理模块, 该模块能够有效地从深度信息中还原被树木和阴影遮挡的建筑轮廓和纹理信息。为了进一步提高网络识别建筑物特征的能力, 通过特征融合模块引入自适应注意力机制, 并实现一个复合损失函数来增强模型对建筑物结构和形态的敏感度。在两个不同的建筑物数据集上进行实验, 结果表明, 该网络的平均交并比(mIoU)分别达到93.60%和96.60%, F1分数、准确率也分别达到94.90%、94.42%和95.90%、96.42%。实验数据显示, 所提网络可以恢复被遮挡信息同时提高分割精度, 为城市规划、资源管理等应用提供有力支持。

关键词: 对抗网络, 无人机, 建筑物分割, 小波变换, 遮挡恢复

Abstract:

Building extraction in Unmanned Aerial Vehicle (UAV) images faces two challenges. First, it is easily blocked by trees and shadows, resulting in segmentation errors. Second, existing methods often ignore the building morphology and multi-resolution information. In the field of segmentation, buildings often encounter occlusion by trees and shadows, and the demand for precise segmentation increases in complex scenes. To address these challenges, this study introduces a dual-channel parallel generator strategy within the framework of adversarial networks. The first channel is based on a morphology-driven wavelet transformation that focuses on capturing the morphological attributes of buildings, including their contours and structural features. The other channel is based on DeepLabv3+ and handles the textural complexity of buildings, including surface textures and details. This design enables the network to comprehend building features from multiple perspectives. Furthermore, to address occlusion issues, this study proposes an occlusion-aware preprocessing module that effectively restores the contours and textural information of buildings occluded by trees and shadows from depth information. An adaptive attention mechanism is introduced through a feature fusion module to enhance the ability of the network to recognize building features. Additionally, a composite loss function is implemented to augment the sensitivity of the model to building structures and forms. The mean Intersection over Union (mIoU), F1-score, and accuracy of the proposed network on two different building datasets are 93.60% and 96.60%; 94.90% and 94.42%; and 95.90% and 96.42%, respectively. Experimental results demonstrate that the proposed network not only restores occluded information but also significantly improves segmentation accuracy, providing robust support for applications such as urban planning and resource management.

Key words: adversarial network, Unmanned Aerial Vehicle(UAV), building segmentation, wavelet transform, occlusion recovery