DiffRefiner: Uncertainty-Aware Diffusion-Driven Latent Correction for Robust Spectrogram Analysis

Hoai-Thuong Truong

doi:10.1109/lcomm.2026.3733310

DiffRefiner: Uncertainty-Aware Diffusion-Driven Latent Correction for Robust Spectrogram Analysis

2026-09-11

0

PRE

AI

DOI：10.1109/lcomm.2026.3733310

Original

Original request for help

Save

Abstract

En 中文

Wireless spectrogram segmentation in complex radio frequency (RF) environments is hindered by severe signal degradation and the coexistence of Long Term Evolution (LTE), 5G New Radio (NR), and radar signals, which induce fragmented spectral signatures and blurred boundaries. Conventional single-pass, deterministic architectures are fundamentally limited by their inability to rectify these structural inconsistencies once they are introduced during the initial encoding stage. To address these limitations, we propose DiffRefiner, a lightweight encoder–latent refinement–decoder framework that reformulates segmentation as an iterative diffusion-driven latent correction process. Treating the encoder output as an initial semantic hypothesis, the model performs refinement within a compact latent space, facilitating targeted correction of structural errors while preserving global semantics. The resulting correction signals are subsequently projected back to the feature domain and selectively integrated into deeper representations via an uncertainty-aware feedback mechanism to enhance robustness in ambiguous regions. Experimental evaluations demonstrate that DiffRefiner achieves 96.98% mean accuracy and 94.86% mIoU with only 4.05 M parameters and 5.46 ms inference latency, establishing a superior accuracy–efficiency Pareto frontier for high-fidelity spectrum monitoring under interference.

Keywords:

Semantic segmentation

diffusion-based refinement

latent diffusion

lightweight segmentation

DiffRefiner: Uncertainty-Aware Diffusion-Driven Latent Correction for Robust Spectrogram Analysis

Abstract

Journal

Organization

Cited Papers