Maritime-Aware Image Deblurring: Benchmark, Information Limit, and a Lightweight Directional Context Module
IEEE Access · 2026 · 14 · 169109-169123
DOI: 10.1109/access.2026.3748712 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Wonjong Lee und 4 weiteren Autor:innen. Die Arbeit nennt 41 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Korea Research Institute for Defense Technology Planning and Advancement (KRIT) Grant funded by Korean Government through the Defense Acquisition Program Admin….
Abstract (Originalsprache)
Electro-optical maritime surveillance imagery is frequently degraded by motion blur from platform vibration and object motion. Most deblurring networks are scene-agnostic, however, and are neither designed for nor evaluated on the structure of maritime scenes, which form near-horizontal bands of sky, horizon, and sea around a small, high-value target: the vessel. This paper makes four contributions. First, we construct MAROB-2500, a maritime deblurring benchmark whose blur is synthesized from measured platform-vibration kernels rather than parametric ones, with the vessel blurred more than the background and scored by a dedicated object-region metric that exposes the small-target failure hidden by overall scores. Second, we show that an on-the-fly severity augmentation over this blur model acts as a strong regularizer, raising a strong baseline by 0.94 dB overall and by 1.61 dB on the information-poor vessel under the same evaluation, about $1.7\times $ the gain on the object as on the whole image. Third, we quantify an information limit on vessel motion blur. It attenuates about 74% of the analyzed frequency band below the 8-bit quantization floor, so that oracle non-blind deconvolution with the true kernel reaches at most 18.5 dB on the vessel, below even the blurred input (19.9 dB) and far below a blind network (23.3 dB), indicating a recoverability limit on the object region that none of our methods exceeds. Fourth, we introduce MSNet (Maritime Scene Network), a lightweight directional-context module inserted into an efficient backbone through a zero-initialized residual, so that each block starts as an identity map. MSNet matches the strong augmented backbone and surpasses every Transformer and multi-scale baseline on overall and background PSNR, at a comparable parameter count and about $10\times $ fewer FLOPs, and restores a full-resolution frame in a single pass. On the object region every method lifts the heavily blurred vessel well above the degraded input, yet all learned methods stay within a narrow band, and MSNet and its backbone are separated by less than the seed-to-seed variation: a single-seed vessel advantage (+ 0.11 dB) does not persist across five seeds, so we claim no object-region gain over the backbone. The information limit explains this directly: scene-aware design keeps the already well-restored background intact and reaches, but cannot exceed, the physical ceiling on the object. The imagery and the measured kernel bank are proprietary, defense-related data and cannot be publicly released; the blur and severity-augmentation models are fully specified here, enabling reproduction of our pipeline on other maritime imagery. The code is publicly available at https://github.com/jk07059-alt/MSNet-MAROB2500
Förderung
- Korea Research Institute for Defense Technology Planning and Advancement (KRIT) Grant funded by Korean Government through the Defense Acquisition Program Admin…
Zitieren
APA
Lee, W., Sin, I., Ha, E., Woo, H., & Kim, S. (2026). Maritime-Aware Image Deblurring: Benchmark, Information Limit, and a Lightweight Directional Context Module. IEEE Access, 14, 169109-169123. https://doi.org/10.1109/access.2026.3748712
Vancouver
Lee W, Sin I, Ha E, Woo H, Kim S. Maritime-Aware Image Deblurring: Benchmark, Information Limit, and a Lightweight Directional Context Module. IEEE Access. 2026;14:169109-169123. doi:10.1109/access.2026.3748712
BibTeX
@article{Lee2026,
author = {Lee, Wonjong and Sin, Ilhan and Ha, Eunjae and Woo, Heejo and Kim, Sungho},
title = {Maritime-Aware Image Deblurring: Benchmark, Information Limit, and a Lightweight Directional Context Module},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {169109--169123},
doi = {10.1109/access.2026.3748712}
}
EUVIMED