DbCrackNet: Dual-branch network for crack segmentation

Abstract

This paper proposes a semantic segmentation network (DbCrackNet) based on convolutional neural network for complex crack detection in concrete surface. The DbCrackNet uses a dual-branch structure based on the encoder-decoder structure and combines a variety of modules such as post-optimized atrous spatial pyramid pooling module, attention mechanism, multi-level supervision mechanism, a new multi-scale and multi-level feature fusion module proposed by us. In the DbCrackNet, different branches use different sizes of input, and the last output of the network is fused by the output of the two branches. It is proved by contrast experiment in our and open crack datasets that DbCrackNet's performance is more superior than Unet, FCN_8S, SegNet, PSPNet and DeepLabV3+ for crack detection. DbCrackNet has smaller loss value, larger mIoU value, larger F1 value, faster convergence rate, more accurate crack detection results and higher anti-interference ability. In conclusion, DbCrackNet can effectively detect cracks in complex environments.