<p>The existing semantic segmentation models generally only use local features to segment the target, and the boundary is unclear due to the loss of pixel-level information in multi-class segmentation tasks, thus reducing the network’s performance. Given the above problems, this paper proposes a semantic segmentation network with dual-path decoding and cascaded multi-level feature interaction(DPD-CMFINet). Firstly, the proposed dual-path decoding structure fully fuses deep semantic features and shallow boundary details through feature fusion and multi-stage upsampling strategy, thereby improving the pixel-level information loss caused by long-distance skip connections. Secondly, the proposed Cascaded multi-level feature interaction module (CMLFIM) makes up for the semantic differences in the feature coding process by utilizing step-by-step feature aggregation and interaction. In addition, to capture multi-scale features from different semantic depths, we construct a hierarchical multi-scale attention module (HMSAM) to enable the network to obtain global context more effectively. Finally, we design a multi-resolution semantic fusion module (MRSFM) to fuse high-resolution features and low-resolution features to obtain more accurate segmentation images. We conducted experiments on multiple datasets to verify the segmentation performance of the proposed model. The segmentation accuracy and mIoU (mean Intersection over Union) of the model on the CamVid (The Cambridge-driving Labeled Video Database) dataset were 93.94% and 80.53%, respectively, demonstrating our proposed approach’s usefulness.</p>