← News

Models & toolsSources: arXiv · 19 Nov 2026

SHAMISA cuts image-quality labeling costs with self-supervised pre-training

Researchers introduce SHAMISA, a non-contrastive self-supervised framework for no-reference image quality assessment that learns from unlabeled distorted images, reducing reliance on costly human perceptual labels.

In plain words

No-reference image quality assessment means judging how good an image looks without comparing it to an undamaged original. Training such models normally needs many expensive human quality ratings. SHAMISA instead pre-trains on unlabeled distorted images, so fewer human scores are needed.

What was announced

The authors describe a compositional distortion engine that generates a wide family of degradations from continuous parameter spaces, grouping images so that exactly one distortion factor varies at a time. This gives fine-grained control over representational similarity during training: images sharing distortion patterns are pulled together in the embedding space, while severity changes produce structured shifts. Dual-source relation graphs combine known degradation profiles with emergent structural affinities throughout pre-training. Unlike earlier approaches with rigid binary similarity constraints, the framework shapes soft, controllable implicit structural associations that are distortion-aware and content-sensitive.

After pre-training, the learned encoder is frozen and paired with a linear regressor. The abstract reports strong overall performance on synthetic, authentic, and cross-dataset benchmarks, improved transfer in the reported cross-dataset comparisons, and favorable qualitative robustness diagnostics. The authors state that the self-supervised pre-training requires no IQA opinion-score annotations or contrastive loss.

What it does not show

  • Evidence is limited to the paper's abstract; the full article's methods, datasets, and benchmark numbers were not available for verification.
  • Performance claims are the authors' own reported results and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.