ML-SPNet: Real-Time Semantic Segmentation Network Based on Multi-Level Serial-Parallel Structure

Real-time semantic segmentation networks suffer from insufficient semantic information extraction, inadequate fusion of information from different branches, and high computational complexity, leading to a decline in segmentation accuracy and slow processing speed. To address these challenges, this paper proposes ML-SPNet, which leverages multi-level serial-parallel architecture to effectively extract contextual information at different scales and enhance the network's understanding of semantic features. Firstly, proposing a lightweight Adaptive Channel Enhancement Module (ACEM), which enhances the representation cap-ability and performance of deep learning models for image semantic information through channel weighting and multi-scale information fusion. Next, introducing the parallel connected channel attention and spatial attention module (P-CSAM) and utilizing channel compression techniques to enhance the network's perception of details and edges while reducing computational costs. Finally, the feature maps from parallel branches are fused in a serial manner to obtain feature outputs with abundant semantic information. Experimental results on the Cityscapes validation dataset demonstrate that ML-SPNet achieves 76.1%mIoU/ 177.3 FPS, exhibiting excellent segmentation accuracy and high processing efficiency. The study validates that the proposed model effectively balances accuracy and real-time performance in the context of autonomous driving.