Surface crack detection of concrete structure based on improved U2-net model
[Objective]In view of the poor continuity and low recognition rate of structural surface cracks with complex back-ground,the crack detection method based on depth learning has the problems of large model parameters.[Methods]This paper constructs a lightweight model U2-Net_Aggregation that aggregates multi-scale information based on the U2-Net framework,which is used to learn fracture characteristics in complex background.By adding jump connections,the model enables each decoding layer to aggregate all shallow coding features above the layer to obtain sufficient feature details and improve the accuracy of crack segmentation;Using depthwise separable convolution(DSC)to improve the ReSidual U-blocks(RSU),a new residual module(RSU-DSC-ECA)is proposed to reduce the problem of increasing model complexity when aggregating multi-scale information.The efficient channel attention(ECA)improves the sensitivity of the model to fracture areas and its anti-interference ability to complex backgrounds.[Results]The ablation experiment was carried out on three sets of fracture data sets.Compared with U2-Net,the improved model(U2-Net_Aggregation)has excellent performance in precision,intersection over union and f1-measure.To verify the model's ability to identify cracks in complex backgrounds,experiments were conducted using concrete structure data collected by UAV,which outperformed FCN,SegNet,U-Net and U2-Net.[Conclusion]The improved model improved by 4.18%,2.97%and 2.03%in recall,intersection over union and f1-measure,respectively,compared to U2-Net,which can quickly and accurately detect cracks with the help of UAV images,providing a new method for structural crack detection.
国家科技期刊平台
NSTL
万方数据
