Semantic segmentation of deep learning has a very broad development prospect in the field of computer vision, but many network models with better segmentation effects take up a lot of memory and take a long time to process a single picture. In response to this problem, we replace the bottleneck unit of the Deeplab V3+ model backbone network (ResNet101) with a 1D non-bottleneck unit, and decompose the convolutional layer of the Atrous Spatial Pyramid Pooling (ASPP) module. The algorithm can greatly reduce the parameter amount of Deeplab V3+ network and accelerate the speed of network inference. Based on the PASCAL VOC 2012 dataset, the experimental results show that the improved network model has faster speed and better segmentation, and takes up less memory space.