AHC-Net: a road crack segmentation network based on dual attention mechanism and multi-feature fusion

To solve the problem of incomplete and inaccurate pavement crack detection, an improved U-Net model based on dual attention mechanism and multi-feature fusion is proposed. Firstly, a new encoding module ACI is designed, which has the feature of multi-scale feature extraction, significantly improves the sensing ability of the damaged area, reduces the background interference, and realizes more accurate segmentation. Secondly, a new decoding module HAD is designed, which avoids the network degradation problem caused by gradient vanishing and the growth of network layers and can retain the most subtle feature information during the decoding process. Finally, convolutional block attention module (CBAM) is introduced in the encoding part to effectively extract global and local detail information, and the criss-cross attention mechanism is also introduced in the decoding part to prevent the loss of marginalized information. The model proposed in this article was tested on the public datasets DeepCrack, CrackSeg478, and AsphaltCrack300, and compared with other advanced methods. The experimental results indicate that this method can detect road cracks more accurately and possesses considerable robustness.