An Efficient Deep Deblurring Technique Using Dark and Bright Channel Priors

Recovering latent sharp image from a blurry one is the target of image deblurring. Learning operation of previous convolutional neural network schemes are based only on content loss function. Our paper presents an efficient blind de-blurring technique that utilizes coarse-to-fine scheme with multi-component loss function to gradually recover the sharp image. The proposed technique applies an end-to-end deep learning strategy to deal with non-uniform motion image blurring cases; it follows a multi-scale network with multi-component loss function. The dark channel prior and bright channel prior are carefully picked to be incorporated into the loss function for network training. The utilized module is based on the scale-recurrent Network (SRN-DeblurNet).