Aiming at the problems of inaccurate segmentation of image target edges and inconsistent segmentation of different types of targets by the DeepLabv3+ model, a semantic segmentation algorithm based on the fusion of class feature attention mechanism is proposed. The algorithm firstly designs a class feature attention module on the encoding end of the DeepLabv3+ model to enhance the correlation between categories, so as to better extract and process semantic information of different categories. Secondly, the multi-level parallel spatial pyramid pooling structure is used to enhance the correlation between spaces so as to better extract contextual information at different scales of images. Finally, at the decoding end, the characteristics of the channel attention module are used to recalibrate the multi-layer fusion features, and the redundant information is suppressed to strengthen the salient features to improve the representation ability of the network. The effectiveness and generalization experiments of the improved model are carried out on the Pascal Voc2012 and Cityscapes datasets, and the average intersection ratios reach 81.34% and 76.27%, respectively.The image edge segmentations in this paper are more detailed, the categories are clearer, and are significantly better than the compared algorithms.