Abstract: The real-time semantic segmentation network does not fully consider the relationship between contextual information and network structure and makes insufficient use of feature information, resulting in rough segmentation. In order to solve the problems, a real-time semantic segmentation network based on multi-level feature reuse (MFRNet) is proposed. A lightweight asymmetric residual attention module (ARAM) was designed to extract rich contextual information and key features. Two efficient feature fusion modules were used to effectively fuse the features of different levels of encoder in a top-down manner, so as to enhance the feature reuse and optimize the segmentation effect. On the Cityscapes and CamVid datasets, the proposed MFRNet achieved 72.6% and 67.3% segmentation precision, 98 FPS and 130 FPS inference speed, respectively. The experimental results show that the proposed algorithm can improve the precision while ensuring real-time segmentation, and it also shows certain advantages compared with recent real-time semantic segmentation algorithms.