Kernel estimation and optimization for image de-blurring using mask construction and super-resolution
- PDF / 1,753,348 Bytes
- 12 Pages / 439.642 x 666.49 pts Page_size
- 35 Downloads / 154 Views
Kernel estimation and optimization for image de-blurring using mask construction and super-resolution Mehwish Iqbal1 · Muhammad Mohsin Riaz2 · Abdul Ghafoor1
· Attiq Ahmad1
Received: 8 June 2019 / Revised: 21 August 2020 / Accepted: 28 August 2020 / © Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract The blur of image is displayed by convolving an image with the blur kernel. Thus, estimating blur kernel is significants of image de-blurring. We aim at obtaining optimized blur kernel of image for de-blurring. Kernel estimation and optimization for de-blurring of image is proposed in this paper. Mask is created for kernel estimation through super pixels and gradient map (generated through illuminant layer). Structural information is extracted through creation of mask through super-pixels, instead of using exemplars and together with the illuminant part of image and gradient map estimates the kernel which is optimized using super-resolution. The proposed method extracts good structural information and edges, hence better de-blurring as compared to state-of-art de-blurring methods. Keywords Super pixels · Kernel optimization · Gradient map · Super-resolution · Edge extraction
1 Introduction In numerous situations, for example, cameras mounted on a moving vehicle or hand-held cameras, it is hard to eliminate camera shake which results in an undesirable blur in the image obtained. De-blurring involves estimating the blur kernel from blur image. Estimation Abdul Ghafoor
[email protected] Mehwish Iqbal [email protected] Muhammad Mohsin Riaz [email protected] Attiq Ahmad [email protected] 1
National University of Sciences and Technology (NUST), Islamabad, Pakistan
2
Centre for Advanced Studies in Telecommunication (CAST), COMSATS, Islamabad, Pakistan
Multimedia Tools and Applications
of blur kernel is very important step in obtaining good quality image. To cater the limitation of obtaining the optimized blur kernel, different kernel estimation techniques for image de-blurring and super-resolution are proposed to improve the image quality.
1.1 Related work Combination of facial component and guided deep convolutional network based on exemplars for kernel estimation result in artifacts around the edges [1]. Convolutional and connected layer estimated acoustic impedance values using unshared weights de-blur images, however, large amount of training data is required [23]. Convolutional neural network using multi-scale features through transfer learning fails to restore the structures from blur image [25]. Encoder-decoder structure based on end to end learning model to refine generative adversarial network, however unable to handle the high frequency features in the image [34]. To learn deep prior, bi-skip network based on generative adversarial network and self-paced learning is an extensive process [38]. Deep neural network based on blind deconvolution cannot handle complex motion blur [19]. Iterative de-blurring based on vectorized step size through updat
Data Loading...
