Abstract:
Low-rank matrix approximation has been successfully used in various image processing
problems like image deblurring, image denoising etc. Previously we studied about low
rank prior of similar patches for image deblurring by combining the properties of the
blurry image and its gradient map. In order to get better kernel estimation, we employed
the weighted nuclear norm minimization method which further enhanced the e ectiveness
of low rank prior by eliminating the inconsiderable edges and ne texture in intermediate
images and preserve the dominant edges of the image. Here we performed both the
quantitative and qualitative analysis for both uniform and non-uniform image deblurring.
We further extend this method of low rank prior using WNNM to jointly perform image
denoising and image deblurring. In this, we show that how to produce a high-quality image
by combining the extracted information from both the denoised image and deblurred
image which cannot be obtained by simply denoising the noisy image or deblurring the
blurred image alone. In this method, we also consider the role residual noise left in the
image even after iteratively denoising the image. In order to further enhance the denoising
process i.e. texture and edges of the image, we utilize the dissimilarity between the
di erent singular values of image patches and then perform image deblurring in order to
obtain the clean image. We compare the result with existing methods and show the improvements
achieved both qualitatively and quantitatively using our method. In the end,
we also implement the dark channel prior technique for both uniform and non-uniform
image deblurring and further extend this method of dark channel prior for image denoising
and compare this method with di erent image denoising and deblurring techniques.
Experiments show that these methods perform favorably against various existing cutting
edge algorithms.
