Séminaire I&M, par Karim Seghouane

Karim Seghouane

Titre : Robust Image Deconvolution Based on α-Divergence

Résumé :

Blind image deconvolution is a fundamental problem in image processing, aiming to recover a sharp image from a blurred and noisy observation without complete knowledge of the degradation process. Using the Gaussian noise assumption, a number of blind image deconvolution algorithms have been proposed in the literature. However, in real world applications the noise is rarely Gaussian and when existing algorithms are used in the presence of violation of this Gaussian assumption, these algorithms provide suboptimal image restoration results. Therefore it is critical to ensure that the image deconvolution algorithm captures this information of violation of the Gaussian assumption.
In this seminar, we present a robust iterative alternating projection method for blind image deconvolution under contaminated Gaussian noise. It’s obtained by using an existing robust measure in information geometry, as an alternating tool to the maximum likelihood blind image restoration. Extensive experiment results comparing the proposed method with state-of-the-art analytical and deep learning techniques under various blur and noise conditions are presented. The results demonstrate the impressive performance, highlighting its robustness and effectiveness in blind image deconvolution with the presence of outliers.
This is a joint work with Dr. Hangfei Zheng and Prof. Djamel Merad.