Focus on hierarchical features: Soft-weighted hierarchical features network

Methods of multi-scale context are commonly used for semantic segmentation. Existing methods suffer from two flaws: (1) Underutilization of backbone-derived multi-scale feature information. (2) Mismatch between small objects and large-scale encodings. To solve the above issues, we propose a novel network to better represent high-level features, named as Soft-weighted Hierarchical Features Network (SWHF-Net) consisting of Semi-atrous Transform Feature Pyramid Module (ST-FPM) and Hierarchical Features Fusion Module (HF2M). Specifically, we propose a hierarchy-driven feature transformation function strat-egy to reconstruct the traditional feature pyramid module as ST-FPM. ST-FPM strengthens the properties of hierarchical features, which is beneficial for extracting the semantic representation of multi-scale objects. Simultaneously, HF2M focuses on the characteristics of features at different hierarchies, and adaptively calculates the attention map of multi-scale objects, greatly improving the efficiency. On Cityscapes, Pascal Context, and ADE20K, we achieve outstanding performance compared to the state-of-the-art methods with fewer computational costs.(c) 2022 Published by Elsevier B.V.

Automated summary

This paper proposes a novel network structure, SWHF-Net, to address the issues in semantic segmentation, including underutilization of backbone-derived features and mismatch between small objects and large-scale encodings. SWHF-Net consists of ST-FPM and HF2M modules, which utilize feature transformation and hierarchical fusion to improve the semantic representation of multi-scale objects and enhance computational efficiency.