Abstarct: Face recognition is a novel subject in the fields of biometric, machine vision and pattern recognition and has wide applications especially in the security systems. As variant factors such as environment lighting, noise and image blurring can more or less (approximately) affect the face recognition performance, thus assessing deblurring methods in order to enhance the recognition accuracy has significant importance. Since inferring the type and characteristics of the point spread function (PSF) of image blurring factor is the main problem in all image enhancement methods, therefore in this thesis, in part of proposed method, by learning the prior knowledge over the training set including the artificially degraded images, inferring the PSF is discussed. According to the proposed method in this thesis, first a set of facial images on the ORL databaxse are artificially blurred and white Gaussian noise of 30 dB is then added. Then we put the features comprised of maximum information related to magnitude of frequency domain components of images degraded by same PSF in a group and trained an MLP neural network over such constructed feature space in learning phase. Then at testing phase, we mapped blurred input facial image with an unknown blurring PSF to the learning stage feature space and extracted the features over the mapped image. Now using trained neural network, we selected the nearest group to this image, amongst learned groups and considered the blurring PSF of this group as the blurring PSF of facial input image. Finally, according to this PSF and using deconvolution, we improved the input image and delivered the improved image to a face recognition system. With proposed method in this thesis, constructing an especial feature space comprised of maximum information related to magnitude of frequency domain components of degraded image, we have enhanced the PSF inference accuracy (inferring accuracy more than 80% in noise condition) and face recognition system accuracy (accuracy was improved from 19.833% to 90.837%) by this method. Also, because of using neural network to infer the PSF, running time is reduced by 41.172 percent compared to an examined novel method in this field.