基于三分支网络的实时图像语义分割

Key Points

• The aim is to develop a real-time image semantic segmentation algorithm that balances accuracy and inference speed.
• Developed a tripartite network structure based on PIDNet.
• Used branches for extracting detail, semantic context, and edge information.
• Implemented an efficient pyramid pooling module for multi-scale context extraction.
• Designed lightweight multi-scale channel interaction attention modules to enhance features.
• Combined features from all three branches to achieve final segmentation results.
• Achieved 79.2% mIoU and 88.5 frame/s on the Cityscapes dataset.
• Achieved 80.5% mIoU and 140.1 frame/s on the CamVid dataset.
• Performance significantly surpasses current baseline methods.