Figures (8)  Tables (2)

    • Figure 1. 

      Classification of asphalt pavement crack detection techniques.

    • Figure 2. 

      Overview of digital image processing techniques.

    • Figure 3. 

      Image acquisition using a CCD camera.

    • Figure 4. 

      Different devices for image acquisition.

    • Figure 5. 

      Flowchart of image denoising using filtering.

    • Figure 6. 

      Development of edge detection technology.

    • Figure 7. 

      Evolution history chart of deep learning models.

    • Figure 8. 

      Development process flowchart of object detection technology based on deep learning.

    • Dataset name Description Application in literature
      Massachusetts dataset Contains 1,711 road images and 151 building images, used for road and building extraction Wei et al.[71] optimized CNN models to extract road categories from aerial images. Alshehhi et al.[72] implemented a patch-based CNN model to extract roads and building parts from remote sensing images.
      TerraSAR-X dataset Used for road extraction in SAR images, with 20% for testing and 80% for training Henry et al.[73] used DeepLabV3+ and Deep Residual U-Net to extract road parts from SAR images.
      DeepGlobe dataset Contains 622 test images, 622 validation images, and 4,971 training images Xie et al.[74] applied a new global perception framework based on higher-order spatial information (HsgNet) for road extraction.
      Google Earth dataset Contains 567 test images and 2,213 training images Cheng et al.[75] proposed the CasNet deep learning model to detect road categories and extract road centerlines. Shi et al.[76] implemented GAN models using data augmentation procedures to generate high-resolution segmentation maps.
      DigitalGlobe dataset Collected by DigitalGlobe satellite, contains 6,226 images Zhou et al.[77] introduced the D-LinkNet model for road semantic segmentation in remote sensing images. Doshi[78] used a ResNet-based ensemble model to extract roads from satellite images.

      Table 1. 

      Common online public datasets.

    • Key technology Main contribution Performance index Ref.
      CNN model optimized based on road structure Combined fusion and deconvolution layers to obtain structured output, proposed a road structure-based
      loss function      		 F1 Score: 66.2%; recall: 72.9%; precision: 60.6%; accuracy: 92.4% Wei et al.[71]
      A novel architecture based on FCN, called U-shaped FCN Proposed U-shaped FCN model, data augmentation
      to improve training efficiency      		 F1 Score: 89.6%; recall: 86.8%; precision: 92.5%; accuracy: 95.2% Kestur et al.[99]
      U-Net fully convolutional network First use of U-Net for detecting concrete cracks,
      U-Net is superior to DCNN in robustness,
      effectiveness, and accuracy      		 F1 Score: 90%; recall: 91%; precision 90% Zhenqing Liu[86]
      Technology based on improved deep encoder-decoder neural networks networks Enhanced model, using ELU function and
      LM method to improve overall output accuracy      		 F1 Score: 85.7%; recall: 86.1%; precision: 85.4% Panboonyuen et al.[100]
      CNN-based automatic bridge crack detection model Using an end-to-end model, parallel use of three Atrous convolutions to reduce computational complexity F1 Score: 87.7%; precision: 78.1%; accuracy: 98.4% Xu & Xu[101]

      Table 2. 

      The development of technology in crack detection.