Blind image deblurring using sparse dictionary and orthogonal projections of filtered patches

In this paper, we present a new blind image deblurring technique based on sparse representation and orthogonal projections of filtered patches. First, we use predefined sparse dictionaries to estimate an initial version of the latent image. To estimate the PSF, four stages are performed. First, a set of directional low-pass filters are applied to the input image. These filters greatly reduces the noise while preserving the blur info in the orthogonal direction. Second, an initial kernel estimation is performed for each filtered image using the initial latent image that was produced by sparse representation. Third, discrete radon transform is applied to each estimated kernel on the orthogonal direction to the low-pass filter and the set of projections at different angles are stored. Fourth, an inverse discrete radon transform is applied to the stored projections and the estimated final kernel is produced. Finally, iterative deconvolution approach is performed to estimate the final latent sharp image. Experimental results showed that the best obtained PSNR is 28.654 at 0.054 noise ratio, whereas the best obtained SSIM is 0.9265 at also 0.054 noise ratio.