Abstract: Due to limitations such as mutual interference among various visible defects in concrete bridges, there is an issue of insufficient exploration of deep contextual information within the segmentation network. This paper proposes a three-branch network (TBNet) for semantic segmentation of multiple apparent lesions in concrete bridges. TBNet comprises three-branch encoder, skip connection and decoder. The three-branch encoder decouples feature information, consisting of progressively hierarchical context branch, “wide and large" detail branche, and “narrow and small" semantic branche. Each branch is responsible for extracting contextual information, detailed information, and semantic information, respectively. The guidance aggregation module (GAM) guides the fusion of contextual information through semantic and detailed information, enhancing the deep contextual information mining capability of the network. Simultaneously, the biformer attention mechanism module reduces redundancy in the encoder's output feature information, further improving segmentation performance. Experimental results demonstrate that TBNet outperforms the baseline U-net (UNet) on the concrete bridge surface multi-defect semantic segmentation dataset, with an improvement in mean pixel accuracy (MPA) of 5.92% and an improvement in mean intersection over union (MIOU) of 6.65%.

Key words: bridge disease segmentation, three-branch encoder, feature information decoupling, context information