An efficient approach based on bayesian MAP for video super-resolution

An Efficient Approach Based on Bayesian MAP for

Video Super-Resolution

Cao Bui-Thu

Department of Electronics Technology

Industrial University of Ho Chi Minh City (IUH)

Ho Chi Minh City, Vietnam

E-mail: [email protected]

Tuan Do-Hong

Department of Electrical-Electronics Engineering

Ho Chi Minh University of Technology

Ho Chi Minh City, Vietnam

E-mail: [email protected]

Thuong Le-Tien

Department of Electrical-Electronics Engineering

Ho Chi Minh University of Technology (HCMUT)

Ho Chi Minh City, Vietnam

E-mail: [email protected]

Hoang Nguyen-Duc

Science and Technology Research Center BRAC

Vietnam Television VTV

Ho Chi Minh City, Vietnam

E-mail: [email protected]

Abstract—Multi-frame super-resolution brings out much

potential to reconstruct real high-resolution video sequences.

This potential is achieved based on its capacity to combine

missing information from different input low-resolution frames.

Although there have been many studies in recent decades, super￾resolution problems for real-world video processing still have

many challenges. This is dues to two problems of: how to address

the affecting factors: motion, sampling and noise explicitly and

how to solve them exactly and efficiently. This paper introduces

an efficient approach for video super-resolution by addressing

real motion, sampling and noise models. Based on that, we

proposed a model for receiving a practical video and an efficient

framework to estimate adaptively the motion and noise to

reconstruct the original high-resolution frames. Our system

achieves promising results when compare with other state-of-the￾art in quality and processing time.

Keywords—image super-resolution; video super-resolution,

Bayesian MAP; neural network.

I. INTRODUCTION

The requirement for high definition of image and video

system is always a great need for many applications, such as:

HDTV information, criminal and civil investigation.

Therefore, converting low-resolution (LR) videos into high￾resolution (HR) videos with noise-free has attracted many

interests and been studied for recent decades.

There are two types of video super-resolution (SR)

methods, SR from single frame and SR from multi-frame.

Single-frame SR methods implement the interpolation in

spatial or frequency domain to up-sample the input LR frame.

Then the HR reconstructed image is applied with filtering,

smoothing and reshaping methods to decrease noise and

increase quality of the reconstructed HR image. With multi￾frame methods, the input LR frames are registered to estimate

the motion (also called optical flow) between them. Then

based on the registered parameters, the input LR frames are

ranged in the same coordinate. The image information missed

in the sampling process of input LR frames will be combined

to interpolate the original HR image. Therefore, the multi￾frame SR methods are always more efficient than the single￾frame SR methods, and they have great potential to

reconstruct the original HR images.

After reviewing the current state-of-the-art, we see that

there are factors that affect the quality of a video SR systems.

They are motion of camera while capturing the scene, motion

of the detailed objects in the scene and the resolution of

camera. In general, the motions of objects and cameras are

arbitrary. The accuracy of estimating these factors plays a

main role in affecting the quality of the reconstructed HR

frame.

Some typical studies can be seen in recently. Prendergast

and Nguyen [1] used a spatial-domain interpolation method.

The registration algorithm use minimum mean-squared error

(MMSE) that based on local correlation to estimate the

motion. In this approach, the author focus simultaneously on

motion estimation and interpolation for the HR frames. Shan

et al, [2] proposed a single-frame SR using iterating de￾convolution algorithm for simultaneously de-blurring and

reconstructing the HR frames. Sroubek et al. [3] proposed a

multi-frame SR algorithm by simultaneously estimating the

blur kerner, motion and reconstructing the HR frames.

However their estimation model is not powerful enough to

achieve improvement. Takeda et al. [4] developed a

framework using 3D kernel regression to exploit the spatio￾temporal neighboring relationship for video SR. However,

deblurring is not solved in this method. So Takeda’s method

provides limited of improvements. Izadpanahi and Demirel [5]

used a combination of multi-frame and single-frame SR

method for static blocks and motion blocks, respectively.

However, their approach does not include blur kernel

estimation and deblurring. Recently, Liu and Sun [6] used

Bayesian approach for adaptive video SR via simultaneously

estimating the underlying motion, blur kernel, and noise level

The 2014 International Conference on Advanced Technologies for Communications (ATC'14)

978-1-4799-6956-2/14/$31.00 ©2014 IEEE 522