| 引入梯度极差正则化的TransUNet无人机建筑外墙裂缝检测研究 |
 点此下载全文 |
| 引用本文：刘向阳1，张莹颖2，葛莹3，吕施霖2，张琛2，肖贞贞2，李阳2，徐延2，龚运2，闫星志1.引入梯度极差正则化的TransUNet无人机建筑外墙裂缝检测研究[J].计算技术与自动化,2026,(2):106-112 |
| 摘要点击次数: 50 |
| 全文下载次数: 0 |
|
|
| 中文摘要:无人机建筑外墙裂缝检测可以实时评估高层建筑墙体的健康状态，有助于建筑结构安全性。外墙裂缝特殊形态会忽略细小裂缝的识别和区分，进而影响无人机影像语义分割品质。将裂缝空间位置间的依赖信息引入TransUNet模型，计算模型预测矩阵各方向的梯度极差，通过在模型损失函数中加入梯度极差正则化项，缓解细小裂缝特征丢失的问题，以此提高目标像素的注意力，从而提升细长带状裂缝的分割性能。研究结果表明，融合梯度极差正则化项的TransUNet模型，其像素分类准确率提升了0.19%，交并比IoU系数提升了5.03%，F1指数提升了4.90%。修正的TransUNet模型对细小裂缝分割更为完整，且分割结果中噪声更少，有助于提升细小裂缝分割的准确率。 |
| 中文关键词:建筑外墙 无人机影像 细小裂缝检测 TransUNet模型 梯度极差正则化 |
| |
| Research on TransUNet with Gradient Range Regularization for UAV-based Building Facade Crack Detection |
|
|
| Abstract:The detection of cracks on the exterior walls based on unmanned aerial vehicles can evaluate the health status of high-rise building walls in real time, which helps to ensure the safety of building structures. The special morphology of external wall cracks may overlook the recognition and differentiation of small cracks, thereby affecting the semantic segmentation quality of images. The method introduces the dependency information between the spatial positions of cracks into the TransUNet model, calculates the gradient range in each direction of the model prediction matrix, and alleviates the problem of feature loss of small cracks by adding a gradient range regularization term to the model loss function, thereby improving the attention of target pixels and enhancing the segmentation performance of slender strip cracks. The research results show that the TransUNet model, which integrates gradient range regularization terms, improves pixel classification accuracy by 0.19%, IoU coefficient by 5.03%, and F1 index by 4.90%. The modified TransUNet model provides a more complete segmentation of small cracks and reduces noise in the segmentation results, which helps to improve the accuracy of small crack segmentation. |
| keywords:building exterior walls UAV imaging small crack detection TransUNet model gradient range regularization |
| 查看全文 查看/发表评论 下载pdf阅读器 |
