This paper proposes an image semantic segmentation method based on an improved DeepLabv3+ network to address the DeepLab network’s inability to fully utilize multiscale feature information while ignoring the problem of high-resolution shallow features and the loss of important pixel information due to excessive direct upsampling multiples. First, the multiscale feature information generated by the network is fully utilized and the feature pyramid network is used to effectively fuse high-resolution shallow features. Then, layer-by-layer upsampling is used to improve the image’s pixel information continuity. Finally, in the atrous spatial pyramid pooling module, the standard convolution is replaced with depthwise separable convolution, enhancing the network model’s training efficiency. The experimental results on the semantic segmentation standard data set PASCAL VOC2012 verification set show that, the mean intersection over union of the method can reach 79.97%. It can predict more refined semantic segmentation results compared with the DeepLab network.