Context Enhanced Real-Time Semantic Segmentation Network in Road Scenes

Semantic segmentation in road scenes plays a critical role in enabling autonomous driving systems. However, current semantic segmentation networks based on deep convolutional network architectures fall short in meeting the stringent real-time and high-precision demands of autonomous driving applications. To enhance the practicality of semantic segmentation in road scenarios, this paper introduces an approach leveraging auxiliary hybrid loss strategy to improve the utilization of low-dimensional and edge features. Additionally, we employ an advanced feature reorganization strategy rooted in channel separation to further enhance the network's semantic information perception. To achieve faster inference, a lightweight backbone network structure is adopted. Our method demonstrates promising results when validated on the Cityscapes dataset.