MobileIQA: Exploiting Mobile-level Diverse Opinion Network For No-Reference Image Quality Assessment Using Knowledge Distillation

---

🚀 🚀 🚀 News:

• ✅ October, 2024: We update the training and testing code.
• ✅ August, 2024: We created this repository.

Checklist

TBU

• Code for MobileIQA
• Code for training
• [] Code for testing
• [] Checkpoint

Catalogue

1. Introduction
2. Usage For Training
3. Usage For Testing
4. Results
5. Citation
6. Acknowledgement

Introduction

With the rising demand for high-resolution (HR) images, No-Reference Image Quality Assessment (NR-IQA) gains more attention, as it can ecaluate image quality in real-time on mobile devices and enhance user experience. However, existing NR-IQA methods often resize or crop the HR images into small resolution, which leads to a loss of important details. And most of them are of high computational complexity, which hinders their application on mobile devices due to limited computational resources. To address these challenges, we propose MobileIQA, a novel approach that utilizes lightweight backbones to efficiently assess image quality while preserving image details through high-resolution input. MobileIQA employs the proposed multi-view attention learning (MAL) module to capture diverse opinions, simulating subjective opinions provided by different annotators during the dataset annotation process. The model uses a teacher model to guide the learning of a student model through knowledge distillation. This method significantly reduces computational complexity while maintaining high performance. Experiments demonstrate that MobileIQA outperforms novel IQA methods on evaluation metrics and computational efficiency.

Usage For Training

Preparation

Run the following commands to create the environment:

conda create -n mobileiqa python=3.10
conda activate mobileiqa
pip install torch torchvision torchaudio

Then run the following command to install the other dependency:

pip install -r requirements.txt

---

You can download the LIVEC, BID, SPAQ, KonIQ and UHDIQA datasets from the following download link.

Dataset	Image Number	Score Type	Download Link
LIVEC	1162 images taken on a variety of mobile devices.	MOS	Link
BID	586 real-blur images.	MOS	Link
SPAQ	11,125 images from 66 smartphone images.	MOS	Link
KonIQ	10,073 images selected from public multimedia resources.	MOS	Link
UHDIQA	6,073 4K images.	MOS	Link

Training process

1. You should replace the dataset path in dataset_info.json to your own dataset path.
2. Run the following command to train the MobileVit-IQA (Please review the train.py for more options).

HF_ENDPOINT=https://hf-mirror.com python3 -u train.py --gpu_id 1 --seed 3407 --batch_size 8 --dataset uhdiqa --loss MSE --model MobileVit_IQA --save_path ./Running_Test

Distillation process

Run the following command to train the MobileNet-IQA with the guidance from the teacher model.

HF_ENDPOINT=https://hf-mirror.com python3 -u train.py --gpu_id 1 --seed 3407 --batch_size 8 --dataset uhdiqa --loss MSE --save_path ./Running_Distill --teacher_pkl YOUR_TEACHER_PKL

Usage For Testing

TBU

Citation

If our work is useful to your research, we will be grateful for you to cite our paper:

@misc{chen2024mobileiqaexploitingmobileleveldiverse,
      title={MobileIQA: Exploiting Mobile-level Diverse Opinion Network For No-Reference Image Quality Assessment Using Knowledge Distillation}, 
      author={Zewen Chen and Sunhan Xu and Yun Zeng and Haochen Guo and Jian Guo and Shuai Liu and Juan Wang and Bing Li and Weiming Hu and Dehua Liu and Hesong Li},
      year={2024},
      eprint={2409.01212},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2409.01212}, 
}

Acknowledgement

We sincerely thank the great work HyperIQA and MANIQA. The code structure is partly based on their open repositories.