This paper presents a practical deep-learning-based crack detection model for inspecting concrete pavement, asphalt pavement, and bridge deck cracks. Crack detection is a typical semantic segmentation task; thus, we propose an encoder-decoder structural model with a fully convolutional neural network, namely, PCSN, by referring to SegNet. This model accepts images of arbitrary size as input data and can be trained pixel by pixel. Moreover, VGG16 net is adopted without the top layer as the encoder, and it is initialized with open-source pretrained weights. “Adadelta” is employed as the optimizer and the cross-entropy is used as the loss function. a crack dataset of images containing complex crack textures is constructed by manual pixelwise annotation. Finally, the dataset is fed into PCSN to train and test the network. FCN-8s and MRCNN are also trained with the same dataset, and the experimental results demonstrate that the PCSN outperforms other algorithm on crack detection, additionally, the basic principle of methodological integration is also briefly introduced.

本文提出了一种基于深度学习的实用裂缝检测模型，用于检查混凝土路面，沥青路​​面和桥面裂缝。裂缝检测是典型的语义分割任务。因此，通过参考SegNet，我们提出了具有全卷积神经网络，即PCSN的编码器-解码器结构模型。该模型接受任意大小的图像作为输入数据，并且可以逐像素进行训练。此外，采用VGG16网络作为顶层编码器，并使用开源的预训练权重对其进行初始化。“ Adadelta”被用作优化器，交叉熵被用作损耗函数。通过手动像素注释构建包含复杂裂纹纹理的图像的裂纹数据集。最后，将数据集输入到PCSN中以训练和测试网络。