README.md

metadata

title: TRIQA Image Quality Assessment
emoji: 🖼️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.0.0
app_file: app.py
pinned: false
license: mit
short_description: TRIQA-IQA

TRIQA: Image Quality Assessment

TRIQA combines content-aware and quality-aware features from ConvNeXt models to predict image quality scores on a 1-5 scale.

Features

• Unified Framework: Single interface combining content-aware and quality-aware feature extraction
• ConvNeXt Architecture: Uses state-of-the-art ConvNeXt models for feature extraction
• Multi-scale Processing: Processes images at two scales (original and half-size) for robust feature extraction
• Regression-based Prediction: Uses trained regression models for quality score prediction
• Easy-to-use Interface: Simple web interface for quality assessment

How It Works

1. Preprocessing: Resize image to two scales (original + half-size)
2. Feature Extraction: Extract content and quality features using ConvNeXt models
3. Prediction: Combine features and predict quality score using regression model

Model Files

Download the required model files from Box and place them in the appropriate directories:

Required Files:

• feature_models/convnext_tiny_22k_224.pth - Content-aware model (170MB)
• feature_models/triqa_quality_aware.pth - Quality-aware model (107MB)
• Regression_Models/KonIQ_scaler.save - Feature scaler
• Regression_Models/KonIQ_TRIQA.save - Regression model (111MB)

Box Links:

• Download Model Files - Place in feature_models/ and Regression_Models/ directories

Citation

If you use this code in your research, please cite our paper:

@INPROCEEDINGS{11084443,
  author={Sureddi, Rajesh and Zadtootaghaj, Saman and Barman, Nabajeet and Bovik, Alan C.},
  booktitle={2025 IEEE International Conference on Image Processing (ICIP)}, 
  title={Triqa: Image Quality Assessment by Contrastive Pretraining on Ordered Distortion Triplets}, 
  year={2025},
  volume={},
  number={},
  pages={1744-1749},
  keywords={Image quality;Training;Deep learning;Contrastive learning;Predictive models;Feature extraction;Distortion;Data models;Synthetic data;Image Quality Assessment;Contrastive Learning},
  doi={10.1109/ICIP55913.2025.11084443}}

Paper Links:

• arXiv: https://arxiv.org/pdf/2507.12687
• IEEE Xplore: https://ieeexplore.ieee.org/abstract/document/11084443

License

MIT License