Classification-Based Adaptive Filtering for Multiframe Blind Image Restoration

In this paper, the blind restoration of a scene is investigated, when multiple degraded (blurred and noisy) acquisitions are available. An adaptive filtering technique is proposed, where the distorted images are filtered, classified and then fused based upon the classification decisions. Finite normal-density mixture (FNM) models are used to model the filtered outputs at each iteration. For simplicity, fixed number of Gaussian components (classes) is, initially, considered for each degraded frame and the selection of the optimal number of classes is performed according to the global relative entropy criterion. However, there exist cases where dynamically varying FNM models should be considered, where the optimal number of classes is selected according to the Akaike information criterion. The iterative application of classification and fusion, followed by optimal adaptive filtering, converges to a global enhanced representation of the original scene in only a few iterations. The proposed restoration method does not require knowledge of the point-spread-function support size or exact alignment of the acquired frames. Simulation results on synthetic and real data, using both fixed and dynamically varying FNM models, demonstrate its efficiency under both noisy and noise-free conditions.