基于多分支深度学习的沥青路面多病害检测方法

Multi-distress detection method for asphalt pavements based on multi-branch deep learning

为了准确、快速地识别路面多病害,采用一种基于多分支框架的深度学习方法,提取并融合路面图像的大、小尺度特征,将路面二维图像和三维图像作为网络输入,增强病害特征.采集裂缝、条状修补、块状修补、坑槽、松散等沥青路面病害图像共计10 562张,进行人工标注.结果表明:500次训练后该方法的平均交并比为0.83,准确率和召回率的调和平均数F值为0.90,优于U-net、PSPNet、DeepLabv3+等方法;在单一类别上,对条状修补、坑槽、松散、桥接缝等分割效果最优,对裂缝、块状修补的识别展现出较强的鲁棒性;所提方法的识别效果高于仅使用单一输入或者单一分支的方法.因此,双通道和多分支的设计方法可以显著提升网络对多类别路面病害的识别精度.

To detect multiple pavement distresses accurately and quickly, a deep learning method based on a multi-branching framework was used to extract and fuse large-scale and small-scale features for pavement images. Two dimensional(2D) images and three dimensional(3D) images of pavement were applied as inputs to the network to enhance the distress features. 10 562 images of asphalt pavement distress such as cracks, sealed cracks, patches, potholes, and raveling were collected and annotated manually. The results show that after 500 training epochs, the mean intersection over unit(MIoU) of this method is 0.83 and the average F value is 0.90, outperforming those of U-net, PSPNet, and DeepLabv3 + methods. In a single category, the segmentation results of sealed cracks, potholes, raveling, and bridge joints are optimal, and the recognition of cracks and patches show strong robustness. The recognition effect of the proposed method is better than those of the methods with only a single input or a single branch. Therefore, the double-channel and multi-branch design method can significantly improve the recognition accuracy of network for multiple categories of pavement distress.