Scholarlink

<p>Deploying semantic segmentation on edge devices presents significant challenges due to the inherent trade-off between model complexity and computational efficiency. Although lightweight backbones are available, conventional symmetric feature fusion modules often suffer from parameter redundancy and the phenomenon of “Semantic Dilution”, in which high-level semantic information is compromised by bottom-up texture noise. To address these limitations, this work introduces the Asymmetric Semantic-Preserving Feature Pyramid Network (AS-FPN). In contrast to standard architectures, AS-FPN utilizes an asymmetric topology that structurally separates the highest semantic level from the bottom-up pathway, thereby mitigating noise propagation. To further enhance efficiency, an isotropic dynamic fusion mechanism is proposed that leverages dynamic channel adaptation to avoid unified large channel dimensions, while employing depthwise convolutions to eliminate redundant learnable weights. Additionally, a global residual recursive architecture is introduced, which stacks lightweight fusion modules and applies post-processing attention mechanisms to the top semantic layers. A dedicated long-skip connection traverses the entire module, ensuring stable gradient flow and maintaining the integrity of the pre-trained backbone. Experimental results indicate that AS-FPN functions as a universal plug-and-play module, substantially surpassing state-of-the-art baselines on standard benchmarks while incurring minimal parameter overhead. The code will be available at <a href="https://github.com/jiaweipan997/AS-FPN">https://github.com/jiaweipan997/AS-FPN</a>.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

AS-FPN: an asymmetric semantic-preserving feature pyramid network for efficient semantic segmentation

  • 期刊论文

  • 发表日期：

    2026年7月11日

查看全文
  • Jiawei Pan,
  • Deyu Zeng,
  • Zongze Wu,
  • Yanyun Qu,
  • Weixiang Liu

摘要

Deploying semantic segmentation on edge devices presents significant challenges due to the inherent trade-off between model complexity and computational efficiency. Although lightweight backbones are available, conventional symmetric feature fusion modules often suffer from parameter redundancy and the phenomenon of “Semantic Dilution”, in which high-level semantic information is compromised by bottom-up texture noise. To address these limitations, this work introduces the Asymmetric Semantic-Preserving Feature Pyramid Network (AS-FPN). In contrast to standard architectures, AS-FPN utilizes an asymmetric topology that structurally separates the highest semantic level from the bottom-up pathway, thereby mitigating noise propagation. To further enhance efficiency, an isotropic dynamic fusion mechanism is proposed that leverages dynamic channel adaptation to avoid unified large channel dimensions, while employing depthwise convolutions to eliminate redundant learnable weights. Additionally, a global residual recursive architecture is introduced, which stacks lightweight fusion modules and applies post-processing attention mechanisms to the top semantic layers. A dedicated long-skip connection traverses the entire module, ensuring stable gradient flow and maintaining the integrity of the pre-trained backbone. Experimental results indicate that AS-FPN functions as a universal plug-and-play module, substantially surpassing state-of-the-art baselines on standard benchmarks while incurring minimal parameter overhead. The code will be available at https://github.com/jiaweipan997/AS-FPN.