Abstract: Aiming at the problems of insufficient precision and robustness of the existing crack detection algorithms in complex environments,a new model full U network is proposed based on the deep learning theory and U-net model.Firstly,the network is constructed based on the U-net model.Then,an upsampling is performed at every pooling layer to restore the feature map specification before this pooling layer andfuse it with the convolution layer before pooling.Finally,the new feature map is concatenated with the layer after upsampling on the U-net.In order to verify the effectiveness of the algorithm,experiments are performed on the test set.Experimental results show that the average precision of the proposed algorithm can reach 83.48%,the recall rate is 85.08%,and F1 is 84.11%.They are 1.48%,4.68%,3.29% higher than the precision,recall,and F1 in U-net respectively.It shows that in a complex environment,the full U network can still extract complete cracks and ensure the robustness.

Key words: Crack detection, Full U network, U-net