Pavement Crack Recognition Method Based on DCGAN Preprocessing and Residual Dense Attention Network
To establish a type of pavement crack detection model based on deep learning convolutional neural network for improving the recognition accuracy for special pavement cracks(e.g.,fuzzy crack and light-colored crack),the pavement crack detection method based on deep convolution generative adversarial network(DCGAN)preprocessing and residual dense attention network(RD AN)was proposed.In the data preprocessing module,guided by the adversarial learning concept,the method preserves high-resolution crack features while suppressing the influence of background noise.The fuzzy cracks can be enhanced to clear cracks.The light-colored cracks can be reconstructed to dark cracks.The training and recognition effect of subsequent models can be improved.In the model training module,a wholly new RDAN for pavement crack recognition was proposed,which was developed through the fusion and improvement of traditional ResNet and DenseNet structures.The layer-to-layer characteristics transfer was synchronized between the element dimension and channel dimension to improve the degradation problem during training of deep convolutional neural networks.The adaptive feature optimization was realized in 2 dimensions of space and channel by introducing the convolutional block attention module to further improve the crack feature extraction ability and suppress the background noise interference in the actual complex road background.The result indicates that the pavement crack recognition method based on DCGAN preprocessing and residual dense attention network has achieved the optimal recognition effect on the CiCS50000 real engineering dataset and multiple public datasets.The accuracy,precision,recall rate and Dice coefficient of the CiCS50000 dataset reach 97.51％,87.05％,83.36％,81.02 respectively.
NETL
NSTL
万方数据
