Comparison of DeepLab v3+ base networks for crack segmentation under limited computational resources

Accurate crack segmentation plays a crucial role in infrastructure assessment and preventive maintenance. This research explored the crack segmentation efficacy of DeepLab v3+, a modern and advanced semantic segmentation network with a high performance and reduced computational cost. The performance comparison was investigated of DeepLab v3+ with different base networks, including Inception-ResNet-v2, Xception, ResNet-50, MobileNet-v2, and ResNet-18. The objective of this paper was to recommend the base network and its optimizer of DeepLab v3+ architecture in terms of crack segmentation of structure for structural health assessment and monitoring under limited resources. The optimizer algorithm, mini-batch size, learning rate, and squared gradient decay factor were adjusted to obtain the best model for each base network considering limited resources of graphics processing unit (GPU) for model training. The best results were analyzed in terms of mean accuracy, class accuracy, and weighted IoU whilst taking the model size into account. The recommended models ranked from the most accurate to the smallest in size are DeepLab v3+ network based on ResNet-50 with Adam optimizer, Xception with RMSProp optimizer, ResNet-18 with SGDM optimizer, and MobileNet-v2 with RMSProp optimizer, respectively. The findings assist in choosing a suitable network architecture for specific applications considering the compromise between model size and performance. The results also highlight the feasibility of the network architecture with tested conditions in terms of structural crack segmentation under limited computational resources.