Single image super resolution of blurred natural images using blur kernel estimation combined with super resolution convolution neural network

—Image De-blurring and super-resolution (SR) are computer vision tasks aiming to restore image detail and spatial scale, respectively. Despite the significant improvement in image quality resulting from improvement in optical sensors and general electronics, camera shake blur significantly undermines the quality of hand-held photographs. We evaluated the state-of-the-art super-resolution convolution neural network(SR-CNN) architecture and proposed a new architecture for SR application inspired by SR-CNN combined with De-blurring. This paper focus super resolution of a de-focussed and motion blurred natural images. Unlike most de-blurring methods that attempt to solve an inverse problem through a variational formulation, deblurring method applied in this work directly estimates the blur kernel by modeling statistical irregularities in the power spectrum of blurred natural images. Extensive experiments indicate that the proposed method not only generates remarkably clear HR images, but also achieves compelling results in PSNR, MSE and SSIM quantitatively.


INTRODUCTION
Single image super resolution (SISR) is a well defined problem in computer vision area. It tries to reconstruct a high resolution image from a single low resolution image. It has been a very attractive research topic over the last two decades [1] [2] [3]. Early SISR methods include interpolation such as bicubic interpolation and Lanczos resampling [4], more powerful methods utilizing statistical image priors [4,5] or internal patch recurrence. Since SISR can restore some high frequency details, it has been applied to many practical applications such as medical imaging, satellite imaging and face identification where rich details are greatly desired. But in many applications images are of poor quality since it suffers from degradation like (i) optical blur (ii) image sampling by the CCD array (iii)motion blur. Motion blur in particular introduces significant image degradation. [7] Image Restoration refers to the construction of the original image given its degraded version, when the phenomenon responsible for the degradation is known. In order to deblur a degraded image, its important to know the characteristics of the blurring process [7]. De-blurring method is used to produce the highest quality image by removing the blur from the degraded image as much as possible. Estimating the blur kernel of an image is the first step towards its deblurring. The state-of-the-art non-uniform deblurring methods can generate clear output, but fail to enlarge the spatial resolution. On the other hand, the existing advanced SR methods are hardly capable of processing blurry LR images well, to address these issues, we have used blur kernel estimation method proposed by Goldstein and Fattal et al, which estimates the kernel by modeling statistical irregularities exhibited in the power spectrum of blurred images, After de-blurring, SRCNN applied for the image spatial resolution enhancement.
In view of the discussions above, this paper focus to generate a high resolution image from a blurred low resolution image that is as close as possible to an ideal image.
The rest of the paper is arranged as follows. Section II describes related work of image super resolution of blurred images, In Section II architecture of proposed methodology.
In Section IV, shown data set details and results.

A. Aim
Generate super-resolving high quality image from LR one with degradation.

B. Design
for i from 1 to N outer do

2) SRCNN Module( Super resolution Convolution Neural Network)
In SRCNN, network have 3 sections, patch extraction and representation, non-linear mapping, and reconstruction as shown in the figure below:

a) Patch Extraction and Representation
It is important to know that the low-resolution input is first upscale to the desired size using bi-cubic interpolation before inputting to SRCNN network. where c is number of channels of the image, f1 is the filter size, and n1 is the number of filters, B1 is the n1dimensional bias vector which is used for increasing the degree of freedom by 1.
Here, it is used for mapping low-resolution vector to highresolution vector.

c) Reconstruction
After mapping, we need to reconstruct the image.
Hence, we do once again.

IV. EXPERIMENTAL RESULTS AND DISCUSSION
An extensive experiments on three publicly available data-sets are performed. The training data-set composed of 91 images by Fatma Albluwi et al. [4]. The test data-sets are denoted "Set5" [4], "Set14" [4] and blur data-set [2,3]

1) Availability of data and materials
The test data-sets are denoted "Set5" [4], "Set14" [4] are available from+ kaggle which is common evaluation dataset for Super Resolution of images. contains various images of buildings to animal faces. Blur data set by Amit Goldstein et al. [2]. The training data-set composed of 91 images by Fatma Albluwi et al. [4].

2) Competing interests
The authors declare that they have no competing interests" in this section.

3) Funding
Not applicable 4) Authors' contributions PP designed and coded proposed methodology under the guidance of V K. Manuscript prepared by PP, read and approved by VK.

5) Acknowledgements
Iam greatful to Dr. Vinu Thomas, Dr. Mini M G, research committee members who have given valid suggestions in this work.