In order to improve the accuracy of semantic segmentation of low -illumination images, an edge en hanced semantic segmentation model (EESN) based on RPN is proposed. Firstly, higher-order semantic features of images were learned by using a deep residual network, and the region proposals were quickly generated by RPN. Secondly, a fusion algorithm was designed to fuse the region proposals and eliminate the repeated candidate regions. Finally, low-illumination edges were searched by the edge search box, and their features were enhanced by using the local enhancement algorithm with low distortion. EESN was used in semantic segmentation of Pascal VOC12 dataset and Cityscapes dataset, and 81.2% and 67.6% of pixel intersection-over-union (mIoU) on these two datasets were obtained. The experimental results also demonstrate that EESN has good segmentation performance for images with low illumination edge.