An evolutionary blind image deconvolution algorithm through the pseudo-Wigner distribution Salvador Gabarda, Gabriel Cristo ´bal * Instituto de O ´ ptica ‘‘Daza de Valde ´s’’ (CSIC), Serrano 121, Madrid 28006, Spain Received 3 March 2005; accepted 22 July 2005 Available online 20 October 2005 Abstract This paper describes a new blind deconvolution method implemented by means of an evolutionary algorithm (EA). The EA is designed following a multi-objective optimisation problem approach. The last generation of the EA is assessed by different quality metrics for determining the solution that provides the best performance. It is shown that different restored images can be obtained from a given testing image. The selection of the best result is accomplished though the use of qual- ity metrics. However, the existence of many quality metrics entails a difficult problem for determining the best output. Here, we present a new robust quality metric, based on the use of the local space-frequency information extracted from the Wigner distribution. We empirically compared its performance with other well-known perceptual metrics. In addition to that, a fusion procedure between all candidate restored output images from the EA is also proposed as an alternative to the selection process. The fusion method is also based on the use of this new measure recently developed by the authors with excellent experimental results. Ó 2005 Elsevier Inc. All rights reserved. Keywords: Evolutionary algorithms; Wigner distribution; Image fusion; Image enhancement; Quality assessment 1. Introduction The aim of image restoration was to provide an improved output from degraded input data. Classical linear image restoration has been exhaustively studied in the past [1,2]. There is a more intricate problem known as blind image restoration which has deserved a major research interest in this area. We can define blind image restoration as a process for estimating the true image from a degraded picture, using only partial or even gross assumptions about the degradation process. In many imaging applications, we can consider an observed image g (x,y) as the result of a two-dimensional convolution of a ground truth image f (x,y) with a linear shift-invari- ant blur h (x,y) operator, known as point-spread function (PSF). Assuming an ideal noiseless scenario, we have J. Vis. Commun. Image R. 17 (2006) 1040–1052 www.elsevier.com/locate/jvci 1047-3203/$ - see front matter Ó 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.jvcir.2005.07.005 * Corresponding author. Fax: +34 91 564 5557. E-mail address: gabriel@optica.csic.es (G. Cristo ´ bal).