Unifying Spatial and Frequency Priors in State Space Models for Robust UAV Image Restoration
Year: 2026
Type: article
Abstract: Aerial imaging is indispensable for numerous applications, yet drone-captured images frequently suffer from complex compound degradations—including motion blur, spatially non-uniform haze, and sensor noise—that severely hinder downstream vision tasks. While deep learning has advanced image restoration, cascading single-task models inevitably causes error accumulation. Furthermore, Vision Transformers achieve high accuracy but incur prohibitive quadratic computational costs, limiting their deployment on edge devices. Recently, State Space Models (e.g., Mamba) have offered linear complexity; however, their 1D sequential scanning disrupts local high-frequency spatial topology, leading to texture blurring. Recognizing that UAV degradations exhibit distinct frequency-domain signatures, we propose UniSpec-Mamba, a unified and efficient restoration network that synergizes frequency-domain priors with selective state-space modeling. Specifically, we design a Multi-Scale Spatial Encoder (MSSE) to extract hierarchical spatial representations. We introduce a dual-branch frequency module utilizing Fast Fourier Transform (FFT) and Discrete Wavelet Transform (DWT) to explicitly decouple low-frequency contours from high-frequency textures, harmonized via a cross-domain fusion mechanism. Finally, a Gated Multi-Scale Decoder (MSD) recovers fine details while suppressing residual artifacts. Extensive experiments demonstrate that UniSpec-Mamba achieves state-of-the-art restoration performance across multiple compound degradation scenarios while maintaining near-linear O(NlogN) computational complexity, making it highly suitable for resource-constrained UAV platforms.
Source: Electronics
Author: Qiaoyue Man
Institution: Gachon University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,896
OpenAlex ID: W7234105665
