Mixed-Degradation Image Restoration using Vision-Language Guidance and Multi-Label Semantic Modulation
| dc.contributor.advisor | Beakal Gizachew | |
| dc.contributor.author | Mintesnot Fikir Abate | |
| dc.date.accessioned | 2026-08-04T21:21:14Z | |
| dc.date.available | 2026-08-04T21:21:14Z | |
| dc.date.issued | 2026-02 | |
| dc.description.abstract | Real-world images are rarely degraded by a single factor; instead, multiple corruptions such as low-light, haze, rain, and blur frequently co-exist and interact across various spatial scales. This mixed-degradation setting exposes critical limitations in current all-in-one restoration methods, which often rely on single-label representations and spatially uniform conditioning. Such approaches fail to model non-exclusive mixtures and capture what degradations are present without identifying where they occur, while providing limited diagnostic transparency for model decisions in practical applications. To address these gaps, we propose the Composite Degradation Semantic Modulation Network (CDSM-Net), a novel framework that decouples degradation reasoning from reconstruction. Our architecture integrates three core components: a wavelet-structured backbone (WAVeViM) for frequency separation, a vision–language semantic controller (MDSM) for multi-label prediction, and a self-gated spatial modulation (SGSM) interface for localized feature conditioning. By leveraging a prototype memory bank and routing weights, the MDSM provides an inspectable intermediate degradation signal and generates a global conditioning code that the SGSM translates into pixel-wise modulation without external masks. Experimental results on composite benchmarks and real-world datasets demonstrate the model’s effectiveness. On the CDD-11 benchmark, CDSM-Net achieves an average PSNR of 29.47 dB and SSIM of 0.9883, improving PSNR by +0.74 dB over OneRestore and attaining the highest average structural similarity among compared state-of-the-art methods. Additional blur-aware stress analysis on CDD-13 further supports the model’s robustness under mixed-frequency composite corruption. Evaluations on the real-world LOL-v2 dataset yield an NIQE of 5.1309, outperforming representative low-light baselines and indicating improved generalization under domain shift. Overall, these results establish CDSM-Net as a robust framework with inspectable degradation routing for resolving non-exclusive mixed degradations in diverse imaging environments | |
| dc.identifier.uri | https://etd.aau.edu.et/handle/123456789/8803 | |
| dc.language.iso | en | |
| dc.publisher | Addis Ababa University | |
| dc.subject | All-in-one image restoration | |
| dc.subject | mixed/composite degradation | |
| dc.subject | visionlanguage guidance | |
| dc.subject | spatial feature modulation | |
| dc.subject | wavelet decomposition | |
| dc.subject | state-space models. | |
| dc.title | Mixed-Degradation Image Restoration using Vision-Language Guidance and Multi-Label Semantic Modulation | |
| dc.type | Thesis |