A lightweight semantic segmentation method for concrete bridge surface diseases based on improved DeeplabV3+

Due to the similar features of different diseases and insufficient semantic information of small area diseases in the surface disease image of concrete bridges, the existing semantic segmentation models for identifying surface diseases in concrete bridges suffer from problems such as large number of parameters, insufficient feature extraction, and low segmentation accuracy. Therefore, this paper proposed a lightweight semantic segmentation method for concrete bridge surface diseases based on improved DeeplabV3+. Firstly, the lightweight improved MobileNetV3 was used as the backbone network to reduce the computational complexity of the model. Secondly, the CSF-ASPP (cross scale fusion atrous spatial pyramid pooling) module was designed to expand the receptive field, enable the model to capture more contextual information at different scales and improve its anti-interference ability. Finally, the focal loss function was used to solve the problem of sample imbalance. The experimental results show that the mean intersection over union (mIoU) and mean pixel accuracy (mPA) of the improved DeeplabV3 + reached 75.24% and 84.68%, respectively, which were 3.73% and 4.21% higher than those of DeeplabV3+. The segmentation accuracy for four diseases of spalling, exposed reinforcement rebar, efflorescence, and crack was better than that of DeeplabV3+, and it also achieved better segmentation results compared to other semantic segmentation models. The improved DeeplabV3 + model achieves a parameter size of 6.97 × 10⁶ and an inference speed of 52.64 FPS, demonstrating 90.33% reduction in parameters and 36.22 improvement in FPS compared to the DeeplabV3+. These advancements significantly enhance its suitability for real-time deployment on edge detection devices while maintaining high segmentation accuracy.