Existing semantic segmentation neural networks require a large amount of floating-point operations and have long running times, and the computational power of current mobile devices is still insufficient compared to laboratory conditions. To address this issue, this paper proposes a semantic segmentation algorithm based on multiple activation functions and feature extraction modules. Firstly, ELU is used to replace the RELU of the main network, which provides more accurate and effective feature extraction for important features of input images, making subsequent classification tasks more efficient. Then, a feature extraction module with an added attention mechanism is introduced to enhance the spatial and channel information of important feature maps containing too many details, while reducing redundant feature maps. Experimental results show that the proposed algorithm achieves an average intersection-over-union (mIoU) of 78.55% on the Cityscapes test set with a resolution input of 1024×2048, while the parameter count is only 3.24M and GFLOPs are only 38.35.