Abstract: In this study, we design a dual-path segmentation algorithm with an embedded improved self-attention mechanism and adaptive fusion of multi-scale features to solve the existence of multiscale targets in the scene image semantic segmentation task and the lack of global context information acquisition in the feature extraction network. We use the simple downsampling module with double branches in the spacial path to perform downsampling four times to extract high-resolution edge detail information, allowing the network to segment the object boundary accurately. Next, we embed the context capture and adaptive feature fusion modules in the semantic path to provide rich multiscale high semantic context information for the decoding stages and adopt a category balance strategy to further enhance the segmentation effect. After experimental verification, the model obtain the indicators of the mean intersection over union (MIOU) of the proposed model are 59.4% and 60.1% on the Camvid and Aeroscapes datasets, respectively, and has a good segmentation effect.