To solve the balance between the excessive computational overhead and the real-time inference speed of the existing semantic segmentation models on mobile devices, a Lightweight Symmetrically Balanced Network (LSBNet) for real-time semantic segmentation is proposed. LSBNet achieves its goals through three key modules: the Detail Reinforcement Module (DRM), the Information Complementarity Module (ICM), and the Effective Aggregation Pyramid Pooling Module (EAPPM). DRM is designed to compensate for the loss of spatial detail information in the feature extraction process to improve segmentation accuracy. ICM solves the imbalance problem in the process of fusion of semantic features and detail features, and enhances the detail information in semantic features through channel attention and spatial attention mechanisms. In addition, EAPPM is used to aggregate global multi-scale contextual information. Experimental results show that LSBNet achieves excellent performance of 73.4% mIoU and 73.7% mIoU on Cityscapes and Camvid datasets, respectively. It is worth mentioning that LSBNet contains only 8.7M parameters, which greatly reduces the number of parameters while maintaining high segmentation accuracy compared to most existing advanced models, demonstrating its superiority in lightweight model design.