Facing unrestricted open vocabularies and diverse scenes， present semantic segmentation methods have the problems of insufficient segmentation， insufficient semantic information extraction and long convergence time. Therefore， this paper proposes a semantic segmentation method that combines U-Net improvement model and superpixel optimization. The U-Net improvement model combines the atrous spatial pyramid pooling （ASPP） and the Xception structure. Firstly， the dilated convolutions （DC） is added to the branch network of the ASPP module to form the serial-parallel structure of the module itself， thus enhancing the image feature extraction capability. And the attention channels are added to the Xception module and a large convolution kernel is used to reconstruct the Xception module， thus reducing the amount of data parameters and increasing the convergence rate. On the basis of the above improvements， the image is then subjected to the super pixel segmentation processing. Finally， conditional random fields are used to impose global constraints on the segmentation results to further optimize the semantic information of pixels. The proposed method is verified on the PASCAL VOC 2012 test set and compared with mainstream networks such as DeepLab V3. Experimental results show that the performance accuracy of the proposed method is increased by 2.4%， which proves the effectiveness of the proposed method in adapting to diverse scenes and dealing with the fine semantic segmentation.