Concrete Bridge Crack Semantic Segmentation Method Based on Improved DeepLabV3+

The traditional semantic segmentation models have low accuracy for the segmentation of bridge cracks, which are elongated and continuous objects with complex backgrounds. Additionally, these models often have a large number of parameters, requiring substantial computational resources for operation. To solve these problems, an improved semantic segmentation algorithm, AFSM-Net, is proposed based on DeepLabv3+. Firstly, the lightweight MobileNetV2 is employed as the backbone network to reduce the number of network parameters. Secondly, the atrous spatial pyramid pooling (ASPP) model is improved by proposing a multi-branch cascaded fusion with the atrous spatial pyramid pooling (MCF_ASPP) model. The parallel convolutional branches within the ASPP module are densely connected to reduce semantic information loss and optimize crack pixel extraction capability. Thirdly, a multi-level feature fusion (MFF) model is proposed. It aims to fuse vision transformers and traditional convolutions to improve the continuity of segmentation. Fourthly, a strip attention gate module is proposed to enhance the extraction of crack edge features and alleviate background pixel interference. The experimental results indicate that our proposed algorithm improves the accuracy of crack segmentation. The evaluation metrics, including recall, precision, and mean intersection over union (MIoU), show values of 87.4%, 82.4% and 85.7%,respectively. These values represent enhancements of 10.2%, 1.1% and 3.9% over the DeepLabV3+ network before our improvements. Furthermore, the parameter count has reduced by 18.48M compared to the original network.