🔥🔥🔥 [2024/06/06] ARNIQA is now included in the IQA-PyTorch 
This is the official repository of the paper "ARNIQA: Learning Distortion Manifold for Image Quality Assessment".
Note
If you are interested in IQA, take a look at our new dataset with UHD images and our latest work on CLIP-based opinion-unaware NR-IQA
No-Reference Image Quality Assessment (NR-IQA) aims to develop methods to measure image quality in alignment with human perception without the need for a high-quality reference image. In this work, we propose a self-supervised approach named ARNIQA (leArning distoRtion maNifold for Image Quality Assessment for modeling the image distortion manifold to obtain quality representations in an intrinsic manner. First, we introduce an image degradation model that randomly composes ordered sequences of consecutively applied distortions. In this way, we can synthetically degrade images with a large variety of degradation patterns. Second, we propose to train our model by maximizing the similarity between the representations of patches of different images distorted equally, despite varying content. Therefore, images degraded in the same manner correspond to neighboring positions within the distortion manifold. Finally, we map the image representations to the quality scores with a simple linear regressor, thus without fine-tuning the encoder weights. The experiments show that our approach achieves state-of-the-art performance on several datasets. In addition, ARNIQA demonstrates improved data efficiency, generalization capabilities, and robustness compared to competing methods.
Comparison between our approach and the State of the Art for NR-IQA. While the SotA maximizes the similarity between the representations of crops from the same image, we propose to consider crops from different images degraded equally to learn the image distortion manifold. The t-SNE visualization of the embeddings of the KADID dataset shows that, compared to Re-IQA, ARNIQA yields more discernable clusters for different distortions. In the plots, a higher alpha value corresponds to a stronger degradation intensity.
@inproceedings{agnolucci2024arniqa,
title={ARNIQA: Learning Distortion Manifold for Image Quality Assessment},
author={Agnolucci, Lorenzo and Galteri, Leonardo and Bertini, Marco and Del Bimbo, Alberto},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={189--198},
year={2024}
}Note
If you want to employ ARNIQA just for inference, you can also use it through the IQA-PyTorch
Thanks to torch.hub, you can use our model for inference without the need to clone our repo or install any specific dependencies. By default, ARNIQA computes a quality score in the range [0, 1], where higher is better.
import torch
import torchvision.transforms as transforms
from PIL import Image
# Set the device
device = torch.device("cuda") if torch.cuda.is_available() else "cpu"
# Load the model
model = torch.hub.load(repo_or_dir="miccunifi/ARNIQA", source="github", model="ARNIQA",
regressor_dataset="kadid10k") # You can choose any of the available datasets
model.eval().to(device)
# Define the preprocessing pipeline
preprocess = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])
# Load the full-scale image
img_path = "<path_to_your_image>"
img = Image.open(img_path).convert("RGB")
# Get the half-scale image
img_ds = transforms.Resize((img.size[1] // 2, img.size[0] // 2))(img)
# Preprocess the images
img = preprocess(img).unsqueeze(0).to(device)
img_ds = preprocess(img_ds).unsqueeze(0).to(device)
# NOTE: here, for simplicity, we compute the quality score of the whole image.
# In the paper, we average the scores of the center and four corners crops of the image.
# Compute the quality score
with torch.no_grad(), torch.cuda.amp.autocast():
score = model(img, img_ds, return_embedding=False, scale_score=True)
print(f"Image quality score: {score.item()}")We recommend using the Anaconda package manager to avoid dependency/reproducibility problems. For Linux systems, you can find a conda installation guide here.
- Clone the repository
git clone https://github.com/miccunifi/ARNIQA- Install Python dependencies
conda create -n ARNIQA -y python=3.10
conda activate ARNIQA
cd ARNIQA
chmod +x install_requirements.sh
./install_requirements.shYou need to download the datasets and place them under the same directory data_base_path.
- [LIVE]: Download the Release 2 folder from here and the annotations from here (corresponding to the realigned subjective quality data)
- [CSIQ]: Create a folder containing the source and distorted images from here and the annotations from here.
- TID2013
- KADID10K
- FLIVE
- SPAQ
For each dataset, move the splits folder placed under the datasets directory of our repo under the
corresponding dataset directory under data_base_path.
At the end, the directory structure should look like this:
├── data_base_path
|
| ├── LIVE
| | ├── fastfading
| | ├── gblur
| | ├── jp2k
| | ├── jpeg
| | ├── refimgs
| | ├── splits
| | ├── wn
| | LIVE.txt
|
| ├── CSIQ
| | ├── dst_imgs
| | ├── src_imgs
| | ├── splits
| | CSIQ.txt
|
| ├── TID2013
| | ├── distorted_images
| | ├── reference_images
| | ├── splits
| | mos_with_names.txt
|
| ├── KADID10K
| | ├── images
| | ├── splits
| | dmos.csv
|
| ├── FLIVE
| | ├── database
| | | ├── blur_dataset
| | | ├── EE371R
| | | ├── voc_emotic_ava
| | ├── splits
| | labels_image.csv
|
| ├── SPAQ
| | ├── Annotations
| | ├── splits
| | ├── TestImage
To get the quality score of a single image, run the following command:
python single_image_inference.py --img_path assets/01.png --regressor_dataset kadid10k--img_path Path to the image to be evaluated
--regressor_dataset Dataset used to train the regressor. Options: ["live",
"csiq", "tid2013", "kadid10k", "flive", "spaq", "clive", "koniq10k"]
By default, ARNIQA computes a quality score in the range [0, 1], where higher is better.
Before training, you need to download the pristine images belonging to the KADIS700 dataset. Download the .zip file from here and unzip it. At the end, the directory structure should look like this:
├── data_base_path
|
| ├── KADIS700
| | ├── ref_imgs
|
| ├── LIVE
|
| ├── CSIQ
|
| ├── TID2013
|
| ├── KADID10K
|
| ├── FLIVE
|
| ├── SPAQ
To train our model from scratch, run the following command:
python main.py --config config.yaml--config <str> Path to the configuration file
The configuration file must contain all the parameters needed for training and testing. See config.yaml for more
details on each parameter. You need a W&B account for online logging.
For the training to be successful, you need to specify the following parameters:
experiment_name: <str> # name of the experiment
data_base_path: <str> # path to the base directory containing the datasets
logging.wandb.project: <str> # name of the W&B project
logging.wandb.entity: <str> # name of the W&B entityYou can overwrite all the parameters contained in the config file from the command line. For example:
python main.py --config config.yaml --experiment_name new_experiment --training.data.max_distortions 7 --validation.datasets live csiq --test.grid_search trueAfter training, main.py will run the test with the parameters provided in the config file and log the results,
both offline and online. The encoder weights and the regressors will be under the experiments directory.
To manually test a model, run the following command:
python test.py --config config.yaml --eval_type scratch--config <str> Path to the configuration file
--eval_type <str> Whether to test a model trained from scratch or the one pretrained by the authors of the paper.
Options: ['scratch', 'arniqa']
If eval_type == scratch, the script will test the encoder related to the experiment_name provided in the
config file or from the command line. If eval_type == arniqa, the script will test our pretrained model.
This work was partially supported by the European Commission under European Horizon 2020 Programme, grant number 101004545 - ReInHerit.
All material is made available under Creative Commons BY-NC 4.0. You can use, redistribute, and adapt the material for non-commercial purposes, as long as you give appropriate credit by citing our paper and indicate any changes that you've made.
