---
kind: concept
slug: image-quality-assessment
title: Image quality assessment
summary: >-
  Scoring how good an image looks, either against a reference image (full-reference) or without one (no-reference). Two no-reference cards: a metric trained on human ratings of super-resolution outputs, and LIQE, a CLIP-based model trained on six quality datasets. Trade-off: the cards are validated on different data, so their scores are not comparable.
tags: [evaluation, metric, image-quality, super-resolution, vision-language]
papers: ["1612.05890", "2303.14968"]
related: [structural-similarity, image-super-resolution, vision-language-pretraining, adaptive-loss-weighting, multi-task-learning, random-forest]
status: draft
updated_at: 2026-10-31
---

## Overview

Image quality assessment (IQA) gives a score to an image that predicts how
good people will find it. The classic setting compares the image with an
undistorted reference, which the paper calls full-reference. The other setting
has no reference and must judge the image alone [arxiv:1612.05890 §1] [§2].
Blind (no-reference) IQA is the setting of [arxiv:2303.14968 §1].

The first paper's motivating case is single-image super-resolution (SR), where
a small image is enlarged. The authors argue that full-reference measures such
as MSE, PSNR and SSIM need the high-resolution ground truth, which is often not
available, and that they match human perception poorly on SR outputs
[arxiv:1612.05890 §1] [tab. 1]. The second paper's motivating case is realistic
camera distortion, where it argues that training tricks based on synthetic
distortions or pseudo-labels do not carry over [arxiv:2303.14968 §1].

The invariant shared by the variants below: the score is fitted or designed to
track human ratings of the same kind of images. The first card trains a
regressor on 81,000 human ratings [arxiv:1612.05890 §3] [tab. 4]. The second
trains on pairs of mean opinion scores from the same dataset with a fidelity
loss [arxiv:2303.14968 §3.2] [eq. 4] [eq. 6].

## Variants

Two no-reference cards are available. The first two subsections describe
them. The remaining subsections compare with the baselines that the first
paper reports, as the paper reports them; those baseline numbers are not
checked here.

### Learned no-reference metric from human ratings

The first paper trains a regressor to predict the mean human score of each SR
image from low-level image statistics, without the original image. The
statistics come from three groups: local DCT block statistics, global
steerable-pyramid statistics and spatial singular values of 5×5 patches
[arxiv:1612.05890 §4.1] [§4.2] [§4.3]. Each group feeds a random forest, and
the three forest outputs are combined linearly with least-squares weights
[arxiv:1612.05890 §4.4]. Card:
[Learning a No-Reference Quality Metric for Single-Image Super-Resolution](/papers/1612.05890).
Author-reported Spearman correlation with human scores is 0.931 under 5-fold
cross-validation [arxiv:1612.05890 tab. 7].

### Vision-language no-reference model (LIQE)

LIQE fine-tunes CLIP ViT-B/32 end to end [arxiv:2303.14968 §4.1]. It scores
each image against 495 short text labels, one for each combination of 5 quality
levels, 9 scene categories and 11 distortion types [arxiv:2303.14968 §3.1]. The
quality score is the expected level under the quality marginal of the same
softmax that gives the scene and distortion labels [arxiv:2303.14968 §3.2]
[eq. 1] [eq. 3]. Training combines a pairwise fidelity loss on quality with
scene and distortion losses, and sets their weights with dynamic weight average
[arxiv:2303.14968 §3.2] [eq. 10] [eq. 11]. The model is trained jointly on six
datasets, with 70/10/20 random splits repeated ten times [arxiv:2303.14968 §4.1].
Card: [Blind Image Quality Assessment via Vision-Language Correspondence: A
Multitask Learning Perspective](/papers/2303.14968).

Author-reported median SRCC over the ten splits: 0.970 on LIVE, 0.936 on CSIQ,
0.930 on KADID-10k, 0.875 on BID, 0.904 on CLIVE and 0.919 on KonIQ-10k
[arxiv:2303.14968 tab. 1]. On the cross-dataset setup it reports 0.811 on
TID2013, 0.881 on SPAQ and 0.478 on PIPAL, and says none of the compared methods
performs well on PIPAL [arxiv:2303.14968 §4.3] [tab. 2].

The paper's ablation favours the fine-tuned encoder: a pretrained CLIP with two
quality levels averages 0.693 mean SRCC, against 0.890 for the fine-tuned version
with the same two levels [arxiv:2303.14968 tab. 4]. Adding the scene and
distortion tasks to quality moves the mean SRCC from 0.915 to 0.922
[arxiv:2303.14968 tab. 5]. These margins are small, and the paper reports no
spread for its ablations [arxiv:2303.14968 tab. 4]. The official code differs
from the paper in several training details, one of them the temperature, which
the paper treats as learnable; see the card's Paper vs code section
[arxiv:2303.14968 §3.2].

### Full-reference measures (baselines)

The first paper compares its metric with IFC, SSIM, FSIM and PSNR, which all use
the ground-truth image [arxiv:1612.05890 §5.2]. Overall Spearman under 5-fold
cross-validation is 0.810 for IFC, 0.635 for SSIM, 0.747 for FSIM and 0.604 for
PSNR [arxiv:1612.05890 tab. 7]. The authors note that these are not a
like-for-like comparison, since they see the reference and the proposed metric
does not [arxiv:1612.05890 §5.2]. SSIM itself is covered on its own page:
[Structural similarity](/concepts/structural-similarity).

### No-reference baselines

The first paper also reports six no-reference methods: BRISQUE, BLIINDS, DIVINE,
BIQI, CORNIA and CNNIQA [arxiv:1612.05890 §5.2]. Their overall 5-fold Spearman
is 0.802, 0.853, 0.589, 0.482, 0.919 and 0.904 respectively [arxiv:1612.05890 tab. 7].
The paper says DIVINE and BIQI are reported with their original models, while
the others are retrained on the paper's data [arxiv:1612.05890 §5.2]. The
paper's own text is inconsistent on this point; see the card's Pitfalls. The
authors' method is ahead of CORNIA (0.919) and CNNIQA (0.904) on the main
protocol. On leave-6-reference-images-out the margin over CORNIA is narrow,
0.852 against 0.843 [arxiv:1612.05890 tab. 7] [tab. 8].

### Semi-reference baseline

NSSA is the one semi-reference method in the first paper's comparison
[arxiv:1612.05890 §5.2]. It scores 0.076 overall under 5-fold cross-validation
[arxiv:1612.05890 tab. 7]. The card records that its learned data file is not
public and that its rebuilding is not described, so this number is weakly
supported [arxiv:1612.05890 §5.2] [tab. 7].

## When to use which

| Situation | Use | Why |
|---|---|---|
| A reference image exists and you can trust it | Full-reference measure | The first paper's full-reference baselines use the ground truth, which a no-reference score lacks [arxiv:1612.05890 §5.2]. On its SR set they score 0.604 to 0.810 overall, below its learned metric at 0.931 [tab. 7] |
| No reference, SR outputs from the same kind of degradation as the first paper | The learned no-reference metric from 1612.05890 | It is validated on this setting: 1,620 SR images from 9 methods at 6 scales [arxiv:1612.05890 §3] [tab. 2] [§5] |
| No reference, a different blur or downsampling model | Treat any score with care | The first paper assumes the kernel width is known and fits it per scale. It does not test other blur models [arxiv:1612.05890 §3] [tab. 2]. LIQE is not trained on SR data either [arxiv:2303.14968 §4.1] |
| No reference, photos with realistic camera distortion, and a GPU for training | LIQE | The paper targets this case [arxiv:2303.14968 §1]. It reports median SRCC 0.919 on KonIQ-10k and 0.904 on CLIVE [tab. 1]. Training ran on one NVIDIA RTX 3090 [arxiv:2303.14968 §4.1] |
| Need the score per image to be cheap | Measure both on your own hardware first | The first paper reports 13.31 s per image in MATLAB on the authors' CPU, with 3 forests of 2,000 trees each [arxiv:1612.05890 §5.1] [§5.3]. The card gives no runtime for LIQE. It scores 15 crops per image, each against 495 texts [arxiv:2303.14968 §3.2] [§4.1] |
| Need code you can reuse under a permissive licence | LIQE, as far as the card shows | The card records MIT for the code [Card](/papers/2303.14968). The first card records no code repository |
| Access to human raters for a new domain | Learned metric, retrained on your own ratings | The first paper rated each image with 50 subjects and averaged 40 of them; which 40 is not stated [arxiv:1612.05890 §3] [§5] |

The first paper gives no evidence that its learned metric beats the baselines
on degradations other than the one it was trained on [arxiv:1612.05890 §3] [§5].
The second paper gives no runtime and no SR-only result [arxiv:2303.14968 §4].

## Implementation sketch

Two sketches follow. The first is the scoring path of the learned metric of
[arxiv:1612.05890 §4]; the second is LIQE inference [arxiv:2303.14968 §3.2].
Both are our own versions, not the papers' code. Feature functions are
described in the first card's Method section and are not spelled out here.

```pseudocode
# Score one image with the learned no-reference metric  [arxiv:1612.05890 §4]
# Input:  image (H, W) grey or (H, W, 3); trained forests and linear weights
# Output: scalar on the 0-to-10 scale of the training ratings  [arxiv:1612.05890 §3]
def score_no_reference(image, forests, weights):
    # Three feature groups, each computed on the image alone  [§4.1] [§4.2] [§4.3]
    features = [
        local_dct_statistics(image),
        global_steerable_statistics(image),
        spatial_singular_values(image),
    ]
    # One regression forest per group, averaged over its trees  [§4.4] [eq. 6]
    partial = [forest.predict(f) for forest, f in zip(forests, features)]
    # Linear combination of the three forest outputs  [§4.4] [eq. 7]
    return dot(weights, partial)
```

```pseudocode
# Score one image with LIQE at inference  [arxiv:2303.14968 §3.2]
# Input:  image (H, W, 3); fine-tuned image encoder; 495 text embeddings
# Output: scalar quality on the 1-to-5 scale of the quality levels  [eq. 1]
def score_liqe(image, image_encoder, text_embeds, tau_1):
    # Sub-images of 224 x 224, cropped without resizing; 15 at test  [§3.2] [§4.1]
    crops = crop_224_without_resizing(image, count=15)
    # Cosine of each crop with each text, averaged over crops  [eq. 2]
    logits = mean(cosine(image_encoder(crops), text_embeds), axis=crops)
    # Softmax over the 5 x 9 x 11 joint labels  [eq. 3]
    p_joint = softmax(logits / tau_1)  # (5, 9, 11)
    # Marginal over the quality levels only  [§3.2]
    p_quality = sum_over_scene_and_distortion(p_joint)  # (5,)
    # Expected quality level  [eq. 1]
    return dot(p_quality, [1, 2, 3, 4, 5])
```

## Contradictions

Across the two papers there is no direct numerical conflict, because they use
different data and protocols. The first paper scores its own SR set of 1,620
images with 5-fold cross-validation [arxiv:1612.05890 §5] [tab. 7]. LIQE uses six
IQA datasets with ten random 70/10/20 splits [arxiv:2303.14968 §4.1] [tab. 1].
The two sets of numbers should not be compared directly. LIQE's baselines do
not include the first paper's metric [arxiv:2303.14968 §4.2].

Inside the first paper, the caption of its Table 10 says the baseline columns
are not retrained, yet three of the four equal the values in its Table 9
[arxiv:1612.05890 tab. 9] [tab. 10]. The card treats those non-retrained
baseline numbers as unverified.

## Papers

- [Learning a No-Reference Quality Metric for Single-Image Super-Resolution](/papers/1612.05890): a no-reference metric for SR outputs, trained on 81,000 human ratings of 1,620 SR images; overall 5-fold Spearman 0.931 as reported by the authors.
- [Blind Image Quality Assessment via Vision-Language Correspondence: A Multitask Learning Perspective](/papers/2303.14968): LIQE, a CLIP-based no-reference model that scores images against 495 joint quality, scene and distortion labels; median SRCC 0.970 on LIVE and 0.919 on KonIQ-10k as reported by the authors.