公路工程

为解决现有卷积神经网络识别桥梁细小裂缝时存在精度低、受环境影响大等问题，提出一种融合语义分割和边缘提取的DEGUNet桥梁裂缝识别方法。首先通过改进传统Canny边缘检测算法，计算图像在多个方向上的梯度幅值，以增强模型的几何敏感度；其次构建动态门控模块并集成于U-Net模型各层级跳跃连接处，提升模型获取多层次裂缝边缘特征的能力；最后结合通道-空间注意力机制以残差连接的方式融入模型解码器结构，提高裂缝检测的精度。将所提方法与VGG-16模型和U-Net模型在相同数据集中进行训练测试，结果表明：所提方法裂缝识别精确率为93.76%,较VGG-16模型和U-Net模型分别提升了19.48个百分点和6.31个百分点；裂缝分割平均交并比为72.35%,相比VGG-16模型和U-Net模型分别提高了10.51个百分点和4.32个百分点。另外，在未参与训练的皮山河大桥数据集中，裂缝识别精确率为92.12%、F₁分数91.92%,召回率91.73%,裂缝分割的平均像素精准度为80.69%,平均交并比为87.38%。因此，本研究提出的DEGUNet网络能够准确分割并识别桥梁细小裂缝，同时在复杂性环境下表现出良好的鲁棒性。

In order to solve the problems of low accuracy and great environmental influence when existing convolutional neural networks are used to identify small cracks in bridges, a bridge crack identification method based on DEGUNet integrating semantic segmentation and edge extraction is proposed. Firstly, the traditional Canny edge detection algorithm is improved to calculate the gradient amplitude of the image in multiple directions to enhance the geometric sensitivity of the model. Secondly, a dynamic gating module is constructed and integrated into the jump connections of each level of U-Net to enhance the model's ability to obtain multi-level crack edge features. Finally, the channel-spatial attention mechanism is integrated into the model decoder structure in the form of residual connection to improve the accuracy of crack detection. The proposed method was trained and tested on the same dataset with the VGG-16 model and the U-Net model. The results showed that the crack identification accuracy of the proposed method was 93.76%, which was 19.48 percentage points and 6.31 percentage points higher than those of the VGG-16 model and the U-Net model, respectively.The mean intersection over union of crack segmentation is 72.35%, which was 10.51 percentage points and 4.32 percentage points higher than those of the VGG-16 model and U-Net model, respectively. In addition, in the Pishan River Bridge dataset that did not participate in the training, the accuracy of crack identification was 92.12%, the F₁ score was 91.92%, the recall rate was 91.73%, the mean pixel accuracy of crack segmentation was 80.69%, and the mean intersection over union was 87.38%. The DEGUNet network proposed in this study can accurately segment and identify small cracks in bridges, while showing good robustness in complex environments.