Cracks are the most common road pavement damage. Due to the propagation of cracks, the detection of early cracks has great practical significance. Traditional manual crack detection is extremely time-consuming and labor-intensive. Researchers have turned their attention to automated crack detection. Although automated crack detection has been extensively researched over the past decades, it is still a challenging task due to the intensity inhomogeneity of cracks and complexity of the pavement environment, e.g. To solve these problems, we propose an efficient pavement crack segmentation model based on deep learning. The model uses pre-trained DenseNet121 as an encoder to extract pavement features. Feature Pyramid Attention module fuses features under different pyramid scales and provides precise pixel-attention. The Global Attention Upsample module which is a combination of convolutional neural network and pyramid module acts as a decoder. The sum of Cross-entropy loss and Dice loss is selected as loss function. We use poly policy to tune learning rate. In order to verify the effectiveness of the proposed method, we conduct training and testing on the Crack500 dataset and MCD dataset. Our method achieves a Dice coefficient of 0.7681, an IoU of 0.6235 on the Crack500 dataset and 0.6909, 0.5278 on the MCD dataset. We perform ablation study to verify the effectiveness of the loss function on improving the performance of our model.

Crack plays a critical role in the field of evaluating the quality of concrete structures, which affects the safety, applicability, and durability of the structure. Due to its excellent performance in image processing, the convolutional neural network is becoming the mainstream choice to replace manual crack detection. In this paper, we improve the EfficientNetB0 to realize the detection of concrete surface cracks using the transfer learning method. The model is designed by neural architecture search technology. The weights are pretrained on the ImageNet. Supervised learning uses Adam optimizer to update network parameters. In the testing process, crack images from different locations were used to further test the generalization capability of the model. By comparing the detection results with the MobileNetV2, DenseNet201, and InceptionV3 models, the results show that our model greatly reduces the number of parameters while achieving high accuracy (0.9911) and has good generalization capability. Our model is an efficient detection model, which provides a new option for crack detection in areas with limited computing resources.

Crack is the early form of most pavement defects and has a great negative effect on road service life. Timely detection and maintenance of cracks may minimize the loss caused by it. In this paper, we propose a lightweight crack segmentation model based on a bilateral segmentation network, which achieves a good balance between inference speed and segmentation performance. The model contains two parts: context path and spatial path. The network used in context path is inspired by Xception, which is used to rapidly down-sample the feature map. Spatial path employs three convolutional layers to encode sufficient spatial information. The F1_score and IoU achieved by our model on the Crack500 dataset are 0.8270 and 0.7379, respectively. The proposed model gains superior performance in FPS compared to other four models. In addition, the model is able to process images at 1,024 × 512 pixels in real-time (31.3 FPS). Through the comparison of training time, our model can save 54.04% of the time.

Water leakage is a critical factor reflecting the structural safety of shield tunnels. Computer vision provides new opportunities to overcome the shortcomings of manual visual inspection and realize automatic detection of water leakage regions. In this study, we propose a leakage segmentation model with an encoder–decoder structure. The encoder adopts multi-branch convolutional attention for feature fusion, and the decoder adopts a lightweight design that only contains multi-layer perceptron. Standard convolution in multi-branch is decomposed to two depth-wise strip convolutions to realize lightweight design and extract strip-like features. Extensive ablation and comparative studies were conducted to test model performance. Test results show that our model achieves robust detection of water leakage under strong noise background, reaching an intersection over union of 90.75% with performance-computation trade-off. Consequently, the proposed method can be an effective alternative to the current visual inspection technologies, and provide a nearly automated inspection platform for shield tunnels.