Skip to main content
Back to timeline
arXivSource publication:

Composite restoration stops treating degradations as additive: PROBE probes cross-degradation responses and reaches 30.77 dB average PSNR on CDD-11

Synopsis

Composite image restoration has long split low light, haze, rain and snow into independent factors, yet one degradation behaves differently depending on which others coexist. PROBE represents degradations as probeable coordinate axes, perturbs features bidirectionally along each axis, and contrasts the responses with intensity-matched single-degradation references to quantify non-additive deviations between factors. On CDD-11, PROBE averages 30.77 dB/0.898 SSIM, above CART's 29.95 dB/0.894, and leads on both metrics across all seven composite conditions.

Source-provided article image: Beyond Additive Composition: Interventional Modeling of State-Dependent Cross-Degradation Responses for Composite Image Restoration
arXiv

Interpretation

Treating composite degradations as additive factors is no longer sufficient; a restoration model needs both how severe each degradation is and how one degradation changes another's response. Earlier all-in-one and composite restoration methods described degradation content and composition through scene-level or factor-level descriptions, implicitly treating coexisting degradations as additive or separable. On haze–snow and haze–rain examples from SCD-15, summing the changes each single degradation induces relative to the same clean image leaves a root-mean-square prediction error of 33.0% and 35.7% of the observed composite latent shift, respectively.

PROBE writes the degradation state as coordinates and measures how other coordinates move when features are nudged along one degradation direction, then uses those responses to guide restoration. Semantically anchored degradation axes align the text embedding offset between degraded and clean prompts with visual feature residual directions, and a coordinate head reads each degradation's relative severity along those axes. In a macro-averaged ablation over the 11 CDD-11 conditions, coordinate conditioning alone raises the backbone from 28.98 dB to 30.17 dB, and coordinates, normalized responses and reference-relative deviations together reach 30.77 dB/0.898 SSIM.

Cross responses must have intensity-matched single-degradation references subtracted, so that what remains is the non-additive part caused by multi-degradation coupling. Because every input is probed along all axes, a single degradation already produces cross-coordinate responses, and using the response matrix directly would misrecord that inherent sensitivity as coupling. On 200 CDD-11 low-light–snow test scenes, predicted rain-to-snow responses compared with snow-coordinate changes after adding rain and re-encoding gave 99.2% per-scene sign agreement, with positive within-scene Spearman correlation in 196 of 200 scenes.

Perspective

The demonstrated scope is three composite restoration datasets, CDD-11, AWMM-100K and SCD-15; SCD-15 synthesizes 15 nonempty degradation combinations at three severities from 2000 clean Cityscapes images, trains on the 11 CDD-11 conditions and tests on all 15, holding out rain–snow, low-light–rain–snow, haze–rain–snow and the four-factor combination from training. The intended readers are researchers and engineers working on all-in-one and composite restoration.

Coordinates are explicitly described as relative degrees rather than probabilities or absolute physical quantities; when cross-response predictions are compared with re-encoded coordinate changes, mean within-scene Spearman correlations are 0.537 under mild and 0.564 under stronger darkening, a moderate correlation; on the AWMM-100K rain–snow condition, MoCE-IR reaches 36.50 dB against PROBE's 35.49 dB.