Adaptive Diffusion Based Restoration for Noisy Facial Image Recognition

Identity recognition from corrupted face image remains difficult, since noise can seriously affect the image quality. Thus, it is necessary to enhance facial image before starting the recognition process. In this paper, we extract relevant features from the noisy facial image using PCA decomposition that seperates the small features from the large ones. For restoring these features, we apply an adaptive diffusion method based on eigenratio of the vectors that represent each feature. Therefore, the denoising process is adapted according to region caracteristics where the small features are enhanced by shock of backward diffusion filter and the large features are smoothed with isotropic diffusion.We have used the ORL database and three different types of noise: Gaussian, uniform and salt-pepper. The proposed method tests six classifiers to find the best one. Numerical experiments show that the proposed method gives the best results for the tested noise types in terms of objective metrics, PSNR and SSIM. They show that SVM classifier provides good performance and outperforms other classifiers with the highest accuracy of 97.85%.