Research map / Image generation / Autoregressive and latent generation
Blind Image Quality Assessment Based on Geometric Order Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes QCN, a transformer-based geometric order learning network that sorts image feature vectors and learnable score pivots in an embedding space for blind image quality assessment, estimating scores by nearest-pivot projection.
Problem
Blind image quality assessment requires estimating continuous quality scores without reference images, and existing methods do not explicitly exploit ordering relations and score differences between images at test time.
Approach
Adapts geometric order learning to BIQA with comparison transformers (CTs) that jointly update image feature tokens and learnable score pivots via masked self-attention and cross-attention, trained with order, metric, center and MAE losses; at test time the score is read off by projecting the test feature onto the nearest pivot pair, using auxiliary training images in the same batch.
Key results
- On KonIQ10K QCN reaches SRCC 0.934 and PCC 0.945, improving over the prior best QPT by 0.76% SRCC and 0.43% PCC.
- QCN achieves best results in 9 of 10 in-dataset tests in Table 1 and best in all 6 cross-dataset tests without pre-training.
- Cross-dataset CLIVE-to-KonIQ10K SRCC of 0.800 is 1.55% above the second-best HyperIQA (per the paper's text).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use QCN instead of QPT | BIQA · KonIQ10K | SRCC: 0.927 -> 0.934 | Helped |
| Use QCN instead of QPT | BIQA · BID, CLIVE, KonIQ10K, SPAQ, and FLIVE | PCC: 0.911 -> 0.890; 0.914 -> 0.893; 0.941 -> 0.945; 0.928 -> 0.928; 0.677 -> 0.741 | Mixed |
| Use QCN instead of the strongest baseline | BIQA · BID, CLIVE, KonIQ10K, SPAQ, and FLIVE | SRCC: 0.888 -> 0.892; 0.895 -> 0.875; 0.927 -> 0.934; 0.925 -> 0.923; 0.646 -> 0.644 | Mixed |
| Add FSU and PSU modules (ablation) | QCN · KonIQ10K | SRCC: 0.858 -> 0.934 | Helped |
| Add FSU module (ablation) | QCN · KonIQ10K | SRCC: 0.858 -> 0.930 | Helped |
| Add PSU module (ablation) | QCN · KonIQ10K | SRCC: 0.858 -> 0.916 | Helped |
| Add Lcenter to Lmae (ablation) | QCN · KonIQ10K | SRCC: 0.855 -> 0.860 | Helped |
| Add Lmetric to Lmae and Lcenter (ablation) | QCN · KonIQ10K | SRCC: 0.860 -> 0.929 | Helped |
| Add Lorder to Lmae, Lcenter, and Lmetric (ablation) | QCN · KonIQ10K | SRCC: 0.929 -> 0.934 | Helped |
| Change input image count N from 1 to 6, 12, 18, and 24 (ablation) | QCN · KonIQ10K | SRCC: 0.909; 0.928; 0.931; 0.935; 0.934 | Helped |
| Change CT count T from 1 to 2, 3, 4, and 5 (ablation) | QCN · KonIQ10K | SRCC: 0.928; 0.930; 0.934; 0.933; 0.933 | Mixed |
| Use non-uniform instead of uniform score pivot generation (ablation) | QCN · KonIQ10K and SPAQ | SRCC: 0.929 -> 0.934; 0.914 -> 0.923 | Helped |
Surprising findings
- Without any self-supervised pre-training, QCN beats pre-training methods such as QPT and Re-IQA on several benchmarks and generalizes best across datasets.
Limitations
- Failure cases where QCN focuses on blurred or poorly lit backgrounds, mis-rating compositions judged by annotators
- Original GOL designed for discrete ranks; direct use is weaker for continuous BIQA
- Requires selecting auxiliary training images at test time, so inference depends on training set
Experimental setup
Datasets: BID CLIVE KonIQ10K SPAQ FLIVE ILSVRC2012
Benchmarks: BID CLIVE KonIQ10K SPAQ FLIVE
Metrics: SRCC PCC
Baselines: NIQE ILNIQE BRISQUE BMPRI CNNIQA WaDIQaM-NR PQR SFA DB-CNN HyperIQA PaQ-2-PiQ UNIQUE
Builds on: geometric order learning order learning MUSIQ-style aspect-ratio preservation Lloyd-Max quantization
Closest papers by meaning
- Adaptive Feature Selection for No-Reference Image Quality Assessment by Mitigating Semantic Noise Sensitivity (ICML 2024 · similarity 0.83)
- Integrating Global Context Contrast and Local Sensitivity for Blind Image Quality Assessment (ICML 2024 · similarity 0.82)
- Distilling Spatially-Heterogeneous Distortion Perception for Blind Image Quality Assessment (CVPR 2025 · similarity 0.82)
- Semi-Supervised Blind Image Quality Assessment through Knowledge Distillation and Incremental Learning (AAAI 2024 · similarity 0.82)
- Feature Denoising Diffusion Model for Blind Image Quality Assessment (AAAI 2025 · similarity 0.82)
- Q-Insight: Understanding Image Quality via Visual Reinforcement Learning (NeurIPS 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.