Image Quality Assessment (IQA)
Image Quality Assessment (IQA)
Predicting how technically degraded an image is relative to some notion of fidelity — noise, blur, compression, banding, transmission artifacts. The output is a scalar quality score meant to track the human Mean Opinion Score (MOS). Distinct from Image Aesthetic Assessment (IAA) (beauty, not fidelity) and from the perceptual quality of generative output (realness, not fidelity) — see the contrast table below. The two problems share their shape, metrics, and method arc, which is why later work scores both with one model.
Two taxonomic axes
IQA is cut along two independent axes; every method announces itself by where it sits on both.
Axis 1 — reference availability
- Full-Reference (FR). Pristine original and distorted copy, pixel- aligned; measure perceptual distance between them. PSNR, SSIM (classical); LPIPS, DISTS (learned). Largely solved — SRCC ≈ 0.98 on legacy sets. Used inside codecs and super-resolution where the clean source is in hand.
- Reduced-Reference (RR). Only a compact set of features from the original (side information over a channel). Least-studied branch; narrow use case.
- No-Reference / Blind (NR / BIQA). Only the image. The reference prior must live inside the model — classically as hand-crafted Natural Scene Statistics, now learned from data. The hard, deployable case: every real deployment (phone uploads, web images, generated images) has no reference. This is where the field concentrated.
Axis 2 — distortion origin
- Synthetic. Known operators (JPEG, blur, noise) applied to pristine images at graded levels. Clean design matrix; supports FR and NR. LIVE, TID2013, CSIQ, KADID-10k.
- Authentic / in-the-wild. Real photos with entangled, unlabelled distortion mixtures and no clean reference — NR only. KonIQ-10k, CLIVE, SPAQ, FLIVE.
Why the field left synthetic-only. Models trained on synthetic distortions learned those specific operators and collapsed on real photos (overfit-to-the-degradation-model). The ~2016–2020 move to authentic sets is the central dataset transition in IQA — a method’s year tells you which distortion world it was built for. See IQA / IAA Datasets.
IQA vs IAA vs generative-perceptual quality
Three orthogonal axes, not one scale — an image can score high on any subset.
| Question | Governed by | Reference exists? | |
|---|---|---|---|
| IQA | How degraded? | noise, blur, compression | sometimes (FR/RR); usually not (NR) |
| IAA | How beautiful? | composition, lighting, subject | never |
| Generative-perceptual | How real? | artifact signatures, texture | distributional (FID, LPIPS over sets) |
A film photograph can be low-IQA (grainy) yet high-IAA (a masterpiece); a diffusion sample can be high on both yet obviously fake. The generative axis is measured distributionally (a set of generated vs. a set of real images), so it sits adjacent to this concept rather than inside it.
Method lineage (NR-IQA)
Both IQA and IAA climbed the same five-rung ladder; the shared skeleton lives in IQA / IAA Evaluation Metrics. The IQA-specific detail — architectures, losses, per-dataset SRCC/PLCC, code repos — is survey report R2. The durable takeaway is that each era removes a specific dependency of the previous one (this failure/repair chain, not the SRCC numbers, is the model):
- Handcrafted NSS (2011–2016) — DIIVINE, BLIINDS-II, BRISQUE, NIQE (opinion-unaware — trained on pristine images only, no MOS), CORNIA/HOSA (codebook). Fit generalised-Gaussian parameters of MSCN coefficients, regress with SVR. Still the speed/interpretability baseline. Broke on authentic photos (compound distortion ≠ clean synthetic operators). KonIQ SRCC ≈ 0.66.
- Deep CNN (2014–2020) — CNNIQA/WaDIQaM (patch training, the small-data workaround), then the transfer-learning turn: DBCNN (two-stream bilinear: synthetic-distortion + ImageNet-semantic), HyperIQA (content-adaptive hyper-network generating per-image weights), PaQ-2-PiQ (patch-to-picture, FLIVE paper), MetaIQA (meta-learning). Fixed small-data via ImageNet transfer. Broke on fixed-resolution resize destroying quality cues, local receptive fields. KonIQ ≈ 0.88–0.91.
- Transformer / ViT (2021–2023) — MUSIQ (multi-scale, native-resolution — the era’s key IQA-specific fix), MANIQA (channel attention, won NTIRE 2022, strong on GAN distortion), TReS (relative-ranking + self-consistency), Re-IQA (contrastive content+quality experts). Fixed resize + global perception; needs the big authentic datasets. Broke on still needing MOS labels. KonIQ ≈ 0.91–0.92.
- Vision-language / CLIP (2022–2023) — CLIP-IQA (antonym “Good/Bad photo” prompts, zero-shot ≈ 0.70 KonIQ with no labels), LIQE (multitask CLIP: joint scene + distortion + quality, ≈ 0.92). Fixed the label dependency: ask a pretrained VLM instead of training a regressor. KonIQ ≈ 0.92.
- Multimodal-LLM scoring (2023–2026) — stop training a regressor; teach a pretrained MLLM to rate. Q-Bench diagnosed that general MLLMs (GPT-4V) perceive low-level attributes but score imprecisely; Q-Instruct taught them to describe quality; Q-Align (ICML 2024, on mPLUG-Owl2) teaches the five discrete words excellent…bad → {5…1} and reads a continuous score as the softmax-weighted average over just those level tokens — because an LLM predicts tokens, ordinal words are in-distribution where a numeral is not (the levels-as-tokens trick). Fixed the label dependency’s successor — the bespoke-model-per-task dependency: OneAlign is one model for IQA + IAA + video VQA, at/above task-specific SOTA. Dents R2’s cross-dataset frontier: Q-Align KonIQ→CLIVE 0.860 vs HyperIQA 0.785, CLIP-IQA+ 0.805 — the pretrained world-knowledge generalizes. In-domain KonIQ 0.940/0.941; DeQA-Score (2025) 0.941/0.953. Full detail and the mechanism live in Multimodal-LLM Visual Scoring; survey R4. Still open: score calibration/reproducibility, explanation hallucination, benchmark contamination via web-scale pretraining, and compute (billions of params vs MANIQA’s ≈ 20 MB).
Also FR perceptual metrics (LPIPS, DISTS, PieAPP) — a parallel branch of deep features as a perceptual distance that beats SSIM; reused across the vision field as differentiable training losses and eval for super-resolution / generative work (see R2), not as blind scorers.
Reference ceiling: NR-IQA on KonIQ-10k reaches ≈ 0.94 SRCC/PLCC at the MLLM frontier (Q-Align, DeQA-Score); transformers/CLIP ≈ 0.92; CNNs ≈ 0.88–0.91; NSS ≈ 0.66. Legacy synthetic sets are saturated (≈ 0.97–0.98 SRCC on LIVE). The unsolved part is cross-dataset generalisation: train KonIQ → test CLIVE drops in-domain ≈ 0.91 to ≈ 0.72–0.79 (HyperIQA 0.906 → 0.785). Numbers and their definitions live in IQA / IAA Evaluation Metrics.
See also
- Image Aesthetic Assessment (IAA) — the aesthetic-quality sibling problem.
- Multimodal-LLM Visual Scoring — where the IQA and IAA tracks merge (rung 5).
- IQA / IAA Datasets — the dataset spine (LIVE → KonIQ → FLIVE).
- IQA / IAA Evaluation Metrics — SRCC/PLCC/KRCC/RMSE, EMD, the method arc.
- Mobile photo ML features (Apple vs Samsung) — how in-product “best-shot”/quality analysis actually ships in iPhone/Galaxy (and why neither vendor names an IQA/IAA model).