Deep Convolutional Neural Network to Handle Saturation in Image Deblurring

Deep neural network techniques are being developed to extract sharp edges from fuzzy photos for kernel estimation. The formulation is based on conventional models, but it has limits and is likely to fail when the blurred images contain a large number of outliers, such as saturated regions. For deblurring photos, several neural networks have been constructed that do not address outliers, while non-neural network algorithms have been designed to manage saturation alone. Few smoothing techniques, edge enhancing filters technique are used as an image pre-processing stage before the deconvolution process, but they don't manage saturated pixels, resulting in ringing artifacts and poor deblurring outcomes. In this work, a masking technique is introduced at pre-processing stage to handle saturation effectively in prior to Deep Convolutional Neural Network for image deblurring. The results show the reduced ringing artifacts and considerable improvements with the performance of existing deep deblur networks.