The work aims to propose an image semantic segmentation optimization method based on DeepLabv3+ network, so as to achieve good image semantic segmentation accuracy, reduce the amount of network parameters as much as possible and speed up network training. The backbone network of encoder was added with attention module and more intensive feature pooling module was used to effectively aggregate multi-scale features. The depthwise separable convolution was applied to reduce the computational complexity of the network. According to the comparison test based on CamVid data set, MIoU score of the enhanced network reached 71.03%, and pixel accuracy and other evaluation indexes such as average pixel accuracy slightly improved compared with the original network. Furthermore, parameters of network were reduced by 12%. The Miou score on the test data set of cityscapes was 75.1%. According to the experimental results, the improved network can effectively extract the feature information of image, improve the semantic segmentation accuracy, and reduce the complexity of the model. The proposed network is tested by the urban street scenes, which can provide reference for the future application of driverless technology, and has certain practical significance.