Lightweight Semantic Segmentation of Road Cracks Based on Improved DeepLabV3+

Aiming at the problem that it is difficult to optimize the accuracy and computational efficiency of semantic segmentation model in road crack detection, this paper proposes a lightweight semantic segmentation network GS-DeepLab, based on the DeepLabV3+ framework, adopting GhostNetV2 as the backbone network, compression of the number of parameters to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{2 5. 1 \%}$</tex> of the original model by the decoupling mechanism of the feature map channel, and integration of the hybrid attention module to strengthen the feature focusing in the crack region. The integration of hybrid attention module strengthens the feature focusing in the crack region, which effectively reduces the computational cost of background interference. Meanwhile, the joint Dice-Focal loss function is designed to improve the crack edge segmentation accuracy by combining the robustness of Dice coefficients to category imbalance and the sensitivity of Focal Loss to boundary pixels. Experiments show that the improved model achieves <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{7 3. 9 4 \%}$</tex> mean intersection and merger ratio (mIoU) on complex road images, and the inference speed is increased to 78 FPS, which is 2.6 times higher than the original model, providing a high-precision and low-latency crack detection solution for mobile device deployment.