Dynamic Multi-Scale Feature Enhanced Network for Semantic Segmentation

To address the issues of low precision in multi-class target segmentation and blurred target boundaries in complex application scenarios, we propose a lightweight semantic segmentation model named DynaSeg, based on the SegFormer network. Firstly, a channel attention mechanism is embedded in the encoder layer to guide the network to adaptively identify key channel information and suppress interference from redundant channels in detail feature extraction. Secondly, a multi-stage spatial gated aggregation module is employed to refine global contextual information through multi-stage aggregation, thereby enhancing the dependencies among semantic information. Finally, a dynamic upsampler is introduced to optimize the multi-scale features fusion process. This recovers some of the boundary information lost due to resolution reduction and improves the classification accuracy for pixels near boundaries. Model variants are constructed using the MiT-B0, MiT-B1, and MiT-B2 series of encoders. The best-performing model achieved an mIoU of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\mathbf{8 1. 5 \%}$</tex> on the Cityscapes dataset. Comparative experimental results demonstrate that this model outperforms the original model on the Cityscapes dataset while retaining the lightweight characteristics of the original network.