Evaluating Image Restoration Without Ground Truth

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2470 papers614 conceptsfrozen 2026-10-04

In the laboratory, image restoration is a solved problem of comparison. To test a new denoising or super-resolution algorithm, researchers take a pristine "ground truth" image, artificially degrade it, and then measure how closely the algorithm can reconstruct the original using metrics like Peak Signal-to-Noise Ratio (PSNR).

In the real world, this paradigm collapses. When a satellite captures a hazy landscape or a radiologist scans a patient, there is no "clean" version hidden in a folder for comparison. This "ground truth gap" makes it nearly impossible to use traditional metrics to tell if an AI is recovering genuine detail or simply hallucinating plausible-looking textures. To bridge this gap, a new direction in research focuses on No-Reference (NR) evaluation—learning to judge and guide restoration using only the corrupted data itself.

The Hallucinated Reference

If a pristine reference does not exist, the most intuitive solution is to create one. Kwan-Yee Lin and Guanxiang Wang pioneered this in Hallucinated-IQA, where a generative network "imagines" what a high-quality version of a distorted image should look like.

The system doesn't need the hallucination to be pixel-perfect; it uses the generated image as a "pseudo-reference" to calculate perceptual loss. By comparing the blurry input to this imagined ideal, the model learns to quantify degradation without ever seeing the true original. Similarly, Hongyu Ren and colleagues proposed the Gain of Restoration (GoR) in RAN4IQA. This approach measures quality by looking at the distance between a distorted image and its AI-restored version—essentially asking, "How much better can a generative model make this look?"

The Degraded Input as an Anchor

While hallucinating a reference is powerful, it risks introducing artifacts. A more conservative approach uses the degraded input itself as a spatial anchor. In Learning Conditional Knowledge Distillation for Degraded-Reference Image Quality Assessment, Heliang Zheng and Huan Yang argue that even a noisy or blurry image contains essential geometric and structural information.

Instead of comparing a restored image to a non-existent clean one, they compare it to the original degraded input. This ensures the restoration stays "faithful" to the captured scene. This concept of consistency is further refined by Guy Ohayon in Proxies for Distortion and Consistency, which demonstrates that we can rank restoration algorithms by measuring their distance to a "Minimum Mean Squared Error" (MMSE) estimate—a statistical proxy that doesn't require ground truth.

Leveraging Discrete Priors and Codebooks

A third strategy involves comparing degraded features against a "dictionary" of high-quality components. Ruiqi Wu and Chunle Guo demonstrated this in RIDCP for image dehazing. By using a pre-trained VQGAN codebook, the model matches hazy, low-quality patches to discrete, high-quality "priors" stored in the network's memory.

This shifts the problem from reconstructing pixels to matching features. If a hazy patch of trees matches a high-quality "tree" entry in the codebook, the model can restore the texture with high confidence. This "reference-conditioned" approach, also explored by Zhifei Zhang in SRNTT, allows models to "borrow" textures from external high-resolution images or internal codebooks to fill in the gaps left by degradation.

Synthetic Supervision and Multi-Modal Feedback

The final hurdle in no-reference research is the lack of human labels. Training these models usually requires thousands of subjective "opinion scores," which are expensive to collect. Peng Ye proposed a workaround in Beyond Human Opinion Scores by using other algorithms to generate "synthetic" quality scores, allowing models to learn the general concept of "quality" without human intervention.

More recently, the field has moved toward using Multimodal Large Language Models (MLLMs) as expert evaluators. In SnowMaster, researchers used MLLMs to provide feedback on desnowing tasks. Because these models are trained on vast amounts of visual and textual data, they can act as a "perceptual judge," identifying where a restoration has over-smoothed a texture or left behind artifacts, providing a sophisticated feedback loop that replaces the need for a pristine reference image.

Papers in this version

Showing 60 of 470 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 375 papers
  2. Version 1 · 2026-10-04 · 470 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_no_reference_quality_assessment_no_refer_057e43ab,
  title = {no-reference quality assessment (no-reference image quality assessment)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/no-reference-quality-assessment-no-reference-image-quality-assessment/rdr_057e43abcec147bb8eeeba09cc869804},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_evaluating_image_restoration_without_gro_057e43ab_v1,
  title = {Evaluating Image Restoration Without Ground Truth},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/evaluating-image-restoration-without-ground-truth/rdr_057e43abcec147bb8eeeba09cc869804/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}