Skip to content
Open access

Dual-path adaptive frequency-aware network for SAR image segmentation

Sep 2026 · ITU Journal on Future and Evolving Technologies · 0 citations

Abstract

Synthetic Aperture Radar (SAR) image segmentation aims to recover coherent semantic structures from observations that are often corrupted by speckle noise. This creates a persistent tension for segmentation methods: broad contextual modeling is needed to judge whether local responses belong to a complete target structure, while fine local evidence is still required to recover boundaries, narrow objects, and small regions. We introduce a Dual-Path Adaptive Frequency-Aware Segmentation Network (DPAFNet) designed around this tension. DPAFNet uses a global-local complementary dual-path encoder to decouple long-range structural reasoning from local detail extraction, a hierarchical adaptive representation bridging module to fuse the two streams with reliability-aware branch weights, and a frequency-perceptive skip modulation module to condition high resolution skip features before decoding. Experiments demonstrate that DPAFNet achieves competitive segmentation performance across overlap, agreement, and boundary metrics. Ablation studies further show the proposed encoder design, adaptive bridging strategy, and frequency-aware skip modulation all contribute to the final performance. These results demonstrate the potential of global-local coordination for improving dense prediction in SAR segmentation tasks.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.