EGB: Image Quality Assessment based on Ensemble of Gradient Boosting

EGB is a No-Reference Image Quality Assessment Metric.

Contents

1. Abstract

2. Getting Started

   2.1. Dependencies

   2.2. Usage Demo

   2.3. More Information on the Provided Files

3. Performance Benchmark

4. Citation

5. Contact

Abstract

Multimedia services are constantly trying to deliver better image quality to users. To meet this need, they must have an effective and reliable tool to assess the perceptual image quality. This is particularly true for image restoration (IR) algorithms, where the image quality assessment (IQA) metric plays a key role in the development of these latter. For instance, the recent advances in IR algorithms, which are mainly due to the adoption of generative adversarial network (GAN)-based methods, have clearly shown the need for a reliable IQA metric highly correlated with human judgment. In this paper, we propose an ensemble of gradient boosting (EGB) metric based on selected features similarity and ensemble learning. First, we analyzed the capability of features extracted by different layers of deep convolutional neural network (CNN) to characterize the perceptual quality distance between the reference and distorted/processed images. We observed that a subset of these layers is more relevant to the IQA task. Accordingly, we exploited these selected layers to compute the features similarity, which are then used as input to a regression network to predict the image quality score. The regression network consists of three gradient boosting regression models that are combined to derive the final quality score. Experiments were performed on the perceptual image processing algorithms (PIPAL) dataset, which has been used in the NTIRE 2021 perceptual image quality assessment challenge.

Getting Started

Dependencies

• python 3.7
• numpy 1.19.5
• Tensorflow 2.4.1
• catboost 0.24.4
• xgboost 0.90
• lightgbm 2.2.3 To use the GPU with tensorflow, these packages should be added:
• CUDA 11.0
• cuDNN 8.0.4

Install the dependencies.

pip install numpy==1.19.5
pip install tensorflow==2.4.1
pip install catboost==0.24.4
pip install xgboost==0.90
pip install lightgbm==2.2.3

Usage Demo

To run the test file

python3 test.py path_to_reference_images path_to_distorted_images

For more Informations

$ python3 test.py --help   
$ python3 test.py -h  

More Information on the Provided Files

1. The directory "models" contains one hdh5 file and three sav files: Pretrained_model.h5, xgboost_model.sav, lightgbm_model.sav and catboost_model.sav.

   • The output of the Pretrained model is a (1,1536) vector containing the necessary features, that will be fed to the three regressors.
   • The output of our Regressors are a (3,1) vector containing the quality estimation using each model, the final quality score is the average of the three scores.

2. The test file contains the main function that is used to generate the output.txt file.

   • Inputs :
     • A path to the reference images. The images should be in this format: xxxx.bmp (the xxxx should contains only letters and numbers ).
     • A path to the distorted images. The images should be in this format xxxx_yyyy.bmp (the yyyy can be any caracter or special caracter or a combinaison of both).
   • Outputs:
     • A text file with the name output.txt, containing in each line the name of the distorted image and the quality score.

Performance Benchmark

Performance comparison on validation set of PIPAL dataset.

Metric	Main Score ↑	SROCC ↑	PLCC ↑
PSNR	0.5464	0.2547	0.2916
NQM	0.7621	0.3457	0.4163
UQI	1.0334	0.4858	0.5475
SSIM	0.7383	0.3399	0.3984
MS-SSIM	1.0496	0.4863	0.5632
IFC	1.2703	0.5936	0.6766
VIF	0.9570	0.4334	0.5235
VSNR	0.6962	0.3212	0.3750
RFSIM	0.5700	0.2655	0.3044
GSM	0.8869	0.4181	0.4688
SRSIM	1.2199	0.5658	0.6541
FSIM	1.0277	0.4671	0.5605
FSIMc	1.0265	0.4678	0.5586
VSI	0.9662	0.4500	0.5161
MAD	1.2340	0.6077	0.6262
NIQE	0.1661	0.0643	0.1017
MA	0.4039	0.2005	0.2034
PI	0.3352	0.1690	0.1662
LPIPS-Alex	1.2738	0.6275	0.6462
LPIPS-VGG	1.2385	0.5914	0.6471
PieAPP	1.4034	0.7062	0.6972
WaDIQaM	1.3322	0.6779	0.6543
DISTS	1.3600	0.6742	0.6858
SWD	1.3291	0.6611	0.6680
EGB (Our)	1.5511	0.7758	0.7752

Performance comparison on test set of PIPAL dataset.

Metric	Main Score ↑	SROCC ↑	PLCC ↑
PSNR	0.5262	0.2493	0.2769
NQM	0.7598	0.3644	0.3953
UQI	0.8695	0.4195	0.4500
SSIM	0.7549	0.3613	0.3935
MS-SSIM	0.9624	0.4617	0.5006
IFC	1.0400	0.4851	0.5548
VIF	0.8765	0.3970	0.4794
VSNR	0.7789	0.3682	0.4107
RFSIM	0.6321	0.3037	0.3284
GSM	0.8740	0.4093	0.4646
SRSIM	1.2087	0.5728	0.6359
FSIM	1.0747	0.5038	0.5709
FSIMc	1.0783	0.5057	0.5726
VSI	0.9752	0.4583	0.5168
MAD	1.1237	0.5433	0.5804
NIQE	0.1658	0.0340	0.1317
MA	0.2873	0.1404	0.1468
PI	0.2490	0.1036	0.1454
LPIPS-Alex	1.1368	0.5658	0.5710
LPIPS-VGG	1.2277	0.5947	0.6330
PieAPP	1.2048	0.6074	0.5974
WaDIQaM	1.1012	0.5532	0.5480
DISTS	1.3421	0.6548	0.6873
SWD	1.2584	0.6242	0.6341
EGB (Our)	1.3774	0.7003	0.6771

Citation

We kindly ask you to reference our paper if you find the repo useful to your research:

@inproceedings{hammou2021egb,
  title={Egb: Image quality assessment based on ensemble of gradient boosting},
  author={Hammou, Dounia and Fezza, Sid Ahmed and Hamidouche, Wassim},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={541--549},
  year={2021}
}

Contact

Hammou Dounia , [email protected]

Fezza Sid Ahmed , [email protected]