mamadou-keita/vsmml

[EUVIP 2021] The official repo for Visual Quality and Security Assessment of Perceptually Encrypted Images based on Multi-Output Deep Neural Network (VSMML)

Visual Quality and Security Assessment of Perceptually Encrypted Images based on Multi-output Deep Neural Network (VSMML)

In this work, we propose a blind CNN-based visual security metric for perceptually encrypted images called VSMML metric. Given an encrypted image, our metric predicts two scores simultaneously, which correspond to the visual security (VS) and visual quality (VQ) scores, as illustrated in Figure below.

Experiments

All experiments are carried out on Google Colab on Mac OS and the detailed results are given in the paper.
Two publicly perceptually encrypted image databases are used in our experiments:

• IVC-SelectEncrypt;
• Perceptually Encrypted Image Database (PEID);
  

This table shows the performance comparison on IVC-SelectEncrypt and PEID datasets with visual quality (VQ) and visual security (VS) scores.

Metrics	IVC-SelectEncrypt			PEID (VQ)			PEID (VS)
	SROCC	KROCC	PLCC	SROCC	KROCC	PLCC	SROCC	KROCC	PLCC
MSE	0.916	0.775	0.683	0.811	0.748	0.890	0.800	0.603	0.810
PSNR	0.916	0.775	0.910	0.813	0.646	0.869	0.797	0.613	0.835
SSIM	0.851	0.689	0.718	0.825	0.670	0.843	0.850	0.677	0.829
FSIM	0.975	0.876	0.896	0.890	0.752	0.911	0.858	0.685	0.880
GMSD	0.968	0.849	0.955	0.801	0.646	0.898	0.754	0.578	0.858
MAD	0.963	0.836	0.952	0.890	0.748	0.905	0.885	0.733	0.898
VIF	0.959	0.832	0.955	0.924	0.797	0.968	0.926	0.787	0.945
NIQE	0.731	0.547	0.496	0.459	0.335	0.327	0.524	0.383	0.528
BRISQUE	0.663	0.485	0.685	0.352	0.251	0.339	0.436	0.305	0.459
ESS*	0.957	0.839	0.948	0.816	0.671	0.922	0.771	0.599	0.891
LSS*	0.953	0.823	0.943	0.798	0.628	0.767	0.770	0.591	0.751
LEG*	0.887	0.723	0.898	0.845	0.681	0.900	0.848	0.666	0.882
LFBVS*	0.895	0.726	0.872	0.634	0.486	0.751	0.630	0.466	0.730
NICE*	0.908	0.759	0.631	0.824	0.486	0.651	0.593	0.437	0.617
LE*	0.093	0.072	0.263	0.092	0.079	0.01	0.155	0.113	0.181
NSD*	0.715	0.529	0.592	0.278	0.196	0.385	0.31	0.214	0.371
VSI-Canny*	0.949	0.819	0.95	0.83	0.708	0.941	0.805	0.635	0.882
QETE*	0.894	0.692	0.726	0.825	0.691	0.853	0.813	0.676	0.818
IIBVSI*	0.968	0.848	0.966	0.642	0.451	0.753	0.878	0.719	0.893
TL-VGG16*	0.943	0.798	0.969	0.892	0.743	0.935	0.935	0.788	0.933
VSMML (Our)*	0.9828	0.9018	0.9794	0.9347	0.8065	0.915	0.9617	0.8445	0.9627

Usage

• Training:
  To train the model on another database, refer to the file TRAIN_ON_PEID.ipynb or TRAIN_ON_IVC.ipynb
• Evaluation:
  To evaluate the performance of our model, please refer to the file Test_ON_PEID.ipynb or Test_ON_IVC.ipynb

Pretrained models

We have trained our models on 80% of datasets and you can find them in ./models

Citation

@inproceedings{fezza2021visual,
  title={Visual Quality and Security Assessment of Perceptually Encrypted Images Based on Multi-Output Deep Neural Network},
  author={Fezza, Sid Ahmed and Keita, Mamadou and Hamidouche, Wassim},
  booktitle={2021 9th European Workshop on Visual Information Processing (EUVIP)},
  pages={1--6},
  year={2021},
  organization={IEEE}
}