Lightweight semantic segmentation via hierarchical scale aggregation and weighted squeezed axial attention

Semantic segmentation is a core task in computer vision, which plays a crucial role in applications such as autonomous driving, remote sensing imagery, and medical imaging. However, existing mainstream models often rely on complex architectures with large parameter scales, making them inefficient for deployment on resource-constrained mobile and embedded devices. To address this issue, this paper proposes a lightweight semantic segmentation network based on hierarchical scale aggregation and weighted squeeze axial attention, termed HSAFormer (Hierarchical Scale Axial Transformer). Specifically, HSAFormer first employs a Hierarchical Scale Aggregation (HSA) module to efficiently capture multi-scale contextual information in the early stages of feature extraction. Then, by jointly modeling the semantic path and spatial path, it effectively learns high-level semantic representations and fine-grained spatial details, respectively. In the semantic path, a Weighted Squeeze Axial Attention (WSAA) module is proposed to model global dependencies with low computational cost. Finally, a Multi-scale and Multi-stage Information Fusion Segmentation Head (MMFHead) is designed to integrate features from different stages. By leveraging a multi-channel scale-scaling mechanism, MMFHead adaptively learns feature importance, enhances semantic continuity, and improves feature representation, thereby further boosting segmentation accuracy. Experiments on ADE20K and Cityscapes datasets demonstrate that the proposed method achieves superior performance compared with other lightweight segmentation models, while maintaining relatively low parameters and computational complexity.