DPDNet: A Real-Time Semantic Segmentation Network with Dual Branches and Progressive Decoding

While a series of semantic segmentation models have achieved outstanding accuracy, their practical applications in domains like autonomous driving and robotics are hindered by their high computational costs. Recently, several real-time semantic segmentation approaches have been proposed. These methods are typically designed as progressive decoding or multi-branch networks, which can effectively improve segmentation accuracy. However, utilizing only a progressive decoding structure may not effectively extract fine-grained details compared to using a multibranch structure. Conversely, relying solely on a multi-branch structure may lack semantic information integration at different stages. To address these issues, we propose a novel real-time semantic segmentation model called DPDNet, which combines dual branches with a progressive decoding network structure. DPDNet consists of two branches: an edge branch and a semantic branch. The edge branch employs an edge loss-guided network to extract more detailed information, while the semantic branch is designed with a progressive decoding structure to fuse semantic information at multiple scales. DPDNet achieves a better balance between segmentation accuracy and inference speed. Specifically, it achieves $76.9 \%$ segmentation accuracy on the Cityscapes test set at a high inference speed of 115.3 FPS.