DIQA: No-Reference Image Quality Assessment Library

DIQA is a library for No-Reference Image Quality Assessment (NR-IQA). It provides a unified interface for predicting image quality using multiple state-of-the-art methods and includes a routing mechanism to select the best method for a given image.

Features

• Unified Inference: Simple API to get MOS (Mean Opinion Score) estimates.
• Smart Routing: Uses an XGBoost-based router to dynamically select the most accurate IQA method for each image.
• Training Support: Built-in DIQATrainer to train the router and mapping coefficients on your own datasets.
• Extensible: Easily integrates with pyiqa for underlying feature extraction.

Installation

pip install diqa

Usage

1. Basic Inference

Predict the quality score of an image:

from diqa import DIQA

# Initialize the engine
diqa = DIQA()

# Predict
result = diqa.predict("path/to/image.jpg")

print(f"MOS Estimate: {result['MOS_estimate']}")
print(f"Method Used: {result['selected_method']}")

2. Training on Custom Data

Train the routing model on your own dataset (requires a CSV with image_name and MOS columns):

from diqa import DIQATrainer

trainer = DIQATrainer(
    image_dir="path/to/images",
    mos_csv="path/to/scores.csv",
    output_dir="my_custom_models"
)

# Prepare data (extract features and compute base scores)
trainer.prepare_data()

# Train the router
trainer.train()

License

MIT