Super resolution algorithm implemented on TMS320C5515 eXdsp USBSTICK

SUPER RESOLUTION ALGORITHM IMPLEMENTED ON

TMS320C5515 eZdsp USBSTICK

T. Le-Tien, C. Bui-Thu, H. Nguyen-Duc, L. Tran-Minh, Q. Huynh N.M., D. Asensi Chicano

EEE Department, Hochiminh city University of Technology, HCMUT, Vietnam

268 Ly Thuong Kiet, Dist. 10, HCMC Vietnam, Tel: +84 903 787 989

Abstract:

The paper is aimed to present an implementation of a

super resolution algorithm on the DSP-kit named

TMS320C5515 USB-Stick. We suggest an algorithm

for combining the spatial domain and the frequency

domain image processing associated with prefilters and

postfiters to obtain a good result with a higher PSNR in

the comparison to other common methods for super￾resolution images. The 2010 updated-version of the

DSP USB-stick from Texas Instrument is successfully

used for checking the match between the simulation on

Matlab and the implemented DSP-hardware.

Keywords: Super-resolution image processing,

Gaussian filter, unsharp filter, TMS320C5515USB-stick,

rotation estimation, shifting estimation, nearest

interpolation algorithm.

1. Introduction to super-resolution images

Briefly basic concept, Image super-resolution technique

is a method for creating a higher resolution image from

lower resolution images. It allows overcoming the

limitations about the resolution of digital image systems

without any hardware improvements. Normally, a super￾resolution algorithm is needed a precise alignment of the

low resolution input images for a successful processing

[2.3]. When using any super-resolution technique to

improve the resolution of image, it is necessary required

to be able to distinguish more details in the final image.

By extracting images of a same scene, more information

can be added to the reproduction. (e.g. creating a super￾resolution image from multi-frames).

Fig.1: Idea situation to obtain a superresolution image

Assuming that there are some differences between the

input images which are often caused by the small

camera movements. In an idea situation [3], we could

assume that of four images taken, the seconds to fourth

image have a horizontal, vertical and diagonal shift of

half a pixel compared to the first image. The pixels from

the first image can be interleaved with pixels from the

three other images to obtain a double resolution image.

2. Rotation estimation algorithm.

Assuming f(x,y) is a low resolution referenced image

and g(x,y) is its shifted and rotated version, the

horizontal shifting is x1,h, vertical shifting is x1,v and

rotation angle is θ. Then the relation of two functions f

and g can be written as [2],

))sin()cos(

( , ) ( cos( ) sin( ) ,

,1

,1

v

h

xxy

g x y f x y x



  



  (1)

With a lightly shifted and rotated version of the images

then 2/1)cos(,sin 2   , and the different

equation between g and f can be written as,

)],() 2 (

) 2 (),([),,(

2

,1

2

,1,1,1

yxg

y

fy xx

x

fx yxyxfxxE

v

hvh

 







  

 

  (2)

Then, the minimum of E(x , x ,1,1 vh , ) can be found

based on partial differential with x1,h, x1,v, and θ. The

value of x1,h, x1,v, θ when E gets min values is the

shifting and rotation parameters. Because of using

approximate values on Taylor series, this method is

limited with the small rotation angle which is must be

less than 60

[2].

3. Shifting estimation algorithm.

Assume f 0(x) is a two-dimension continuous signal

and its shifting version 1 xf )( , we have [3]:

f (x)  f (x  x101 )Where 





 





  v

h

x

x

x and 





 





  v

h

x

x

x

,1

,1

1

dxexfF

x T j

x

  2 11 )()( 

 

dxexxf

x T j

x

2  10 )( 

  (3)

x'  x  x1 , then

   '

'2 0 2 1 ')'()( 1

x

x T jx T j dxexfeF  

0 )( 2 1

  Fe

x T j  (4)

 2))(/)(( xjFF 101 T   (5)