PROJECT TITLE :

Graph-Based Blind Image Deblurring From a Single Photograph

ABSTRACT:

Deblurring an image without knowing the blur kernel is an extremely difficult task. It is possible to solve the issue in two ways: A blur kernel can be estimated by looking at the blurry image, and then deconvolving the fuzzy input to get the target image back. By interpreting an image patch as a signal on a weighted graph in this research, we have devised an algorithm for de-blurring a picture. Skeleton images can be used to accurately estimate the blur kernel and have an unusual bi-modal edge weight distribution, which we suggest can be utilised to accurately estimate the blur kernel. Reweighted graph total variation priors are then designed to encourage a bimodal edge weight distribution when the patch is fuzzy. Furthermore, we offer a novel weight function to express RGTV as a graph l1-Laplacian regularizer in the graph frequency domain. Using this approach, we can derive a graph spectral filtering interpretation of the prior with desirable qualities, including resilience to noise and blur, strong piecewise smooth filtering, and sharpening promotion. RGTV's non-convex non-differentiable optimization problem for a blind image deblurring goal is non-convex. An efficient approach for solving for the skeleton image and the blur kernel is designed using the new graph spectral interpretation for the RGTV. A new technique for blind Gaussian deblurring employing graph spectral filtering has been proposed for Gaussian blur. It is now possible to use non-blind picture deblurring methods to restore the target image with the computed blur kernel. Using our approach, we were able to restore latent crisp pictures in a quantitative and qualitatively superior way to the current state of the art.