|
|
--- |
|
|
tags: |
|
|
- image-quality-analysis |
|
|
--- |
|
|
## Note on This Upload |
|
|
|
|
|
The weights included here (`HiRQA.pth`, `HiRQA-S.pth`) are **not modified**—they are simply **re-uploaded from the original authors for easier access and integration within the Hugging Face ecosystem** (e.g., `hf_hub_download`, XetFS, etc.). |
|
|
All credit for the models and methodology goes to the original authors. |
|
|
|
|
|
|
|
|
Refer to original github repo for details: https://github.com/uf-robopi/HiRQA. |
|
|
|
|
|
|
|
|
|
|
|
## Model Variants |
|
|
|
|
|
- **HiRQA (ResNet-50 backbone)** — higher accuracy, suitable for offline evaluation. |
|
|
- **HiRQA-S (ResNet-18 backbone)** — optimized for real-time applications. |
|
|
|
|
|
|
|
|
|
|
|
--- |
|
|
|
|
|
|
|
|
|
|
|
## HiRQA: Hierarchical Ranking and Quality Alignment for Opinion-Unaware Image Quality Assessment |
|
|
|
|
|
**HiRQA** is an opinion-unaware no-reference image quality assessment (NR-IQA) framework that learns a hierarchical, quality-aware embedding space without requiring human opinion scores during training. It introduces three key components: |
|
|
|
|
|
- **Pair-of-Pairs Ranking Loss** — enforces consistent hierarchical relationships between distortions. |
|
|
- **Embedding Distance Consistency Loss** — stabilizes relative quality ordering. |
|
|
- **Contrastive Image–Text Alignment** — improves generalization to real-world distortions via CLIP-based semantic cues. |
|
|
|
|
|
HiRQA requires only a **single distorted image at inference**, and its lightweight variant **HiRQA-S** provides real-time performance. |
