摘要
热桥是建筑围护结构中的薄弱环节，可能导致能源损失、湿气积聚以及建筑构件中霉菌的滋生。为检测大规模建筑存量中的热桥，可采用搭载热成像相机的无人机进行巡检。由于对大规模图像数据集进行人工分析耗时极长，本文致力于探索深度学习方法以实现该过程的自动化。为此，我们聚焦于从更新后的《建筑屋顶热桥数据集》（TBBRv2）中获取的全景无人机图像，该数据集包含926张图像及6,927个标注，涵盖RGB图像、热成像数据与高程信息。我们对比了五种不同神经网络架构的先进模型在有无预训练条件下的表现：MaskRCNN R50、Swin-T Transformer、TridentNet、FSAF，以及一个MaskRCNN R18基线模型。实验结果表明，预训练模型表现尤为突出，其中基于预训练Swin-T Transformer的模型在检测大型热桥时平均召回率（Average Recall）超过50%，展现出良好的应用前景。
基准测试
| 基准 | 方法 | 指标 |
|---|---|---|
| instance-segmentation-on-tbbr | Swin-T (ImageNet-1k pretrain) | Average Recall@IoU:0.5-0.95: 28.0 |
| instance-segmentation-on-tbbr | Mask R-CNN (ResNet-50-FPN) | Average Recall@IoU:0.5-0.95: 20.1 |
| instance-segmentation-on-tbbr | Swin-T | Average Recall@IoU:0.5-0.95: 20.6 |
| instance-segmentation-on-tbbr | Mask R-CNN (ResNet-50-FPN, ImageNet-1k pretrain) | Average Recall@IoU:0.5-0.95: 21.9 |
| object-detection-on-tbbr | Mask R-CNN (ResNet-50-FPN) | Average Recall@IoU:0.5-0.95: 30.8 |
| object-detection-on-tbbr | Mask R-CNN (ResNet-50-FPN, ImageNet-1k pretrain) | Average Recall@IoU:0.5-0.95: 37.0 |
| object-detection-on-tbbr | FSAF (ResNeXt-101, ImageNet-1k pretrain) | Average Recall@IoU:0.5-0.95: 38.0 |
| object-detection-on-tbbr | TridentNet (ResNet-50, ImageNet-1k pretrain) | Average Recall@IoU:0.5-0.95: 30.0 |
| object-detection-on-tbbr | TridentNet (ResNet-50) | Average Recall@IoU:0.5-0.95: 21.5 |
| object-detection-on-tbbr | FSAF (ResNeXt-101) | Average Recall@IoU:0.5-0.95: 24.8 |
| object-detection-on-tbbr | Swin-T (ImageNet-1k pretrain) | Average Recall@IoU:0.5-0.95: 45.4 |