Lightweight Attention Fusion Dual-Branch Feature Network for Semantic Segmentation of Traffic Scenes

Focus on the low computational efficiency of semantic segmentation in traffic scenes, large number of model parameters, and loss of information for small objects, a lightweight attention fusion dual-branch feature network for semantic segmentation is proposed. First, we propose a lightweight convolution-based asymmetric residual module to extract multi-scale information from different receptive fields while decreasing computational and parametric quantities. Secondly, we propose spatial detail extraction modules using shallow wide structures that establish local dependencies for each pixel to preserve detail information. Finally, a lightweight feature fusion module with a bilateral pooling attention strategy is proposed to enhance semantic consistency between different branch features in channel and space dimensions. Experimental results show that the proposed method achieves an accuracy of 77.20% mean IoU at the number of model parameters of 3.04M on Cityscapes data sets. The proposed method demonstrated robustness for small objects and object edges in traffic scene images.