An improved nearest-neighborhood algorithm for efficient supper-resolution images implemented on ARM9 AT91SAM9RL

AN IMPROVED NEAREST-NEIGHBORHOOD ALGORITHM FOR EFFICIENT

SUPPER-RESOLUTION IMAGES IMPLEMENTED ON ARM9 AT91SAM9RL

Thuong Le-Tien, Cuong Nguyen Hung, H Marie Luong*, Cao Bui Thu, Hoang Duc Nguyen

Hochiminh city University of Technology, Vietnam; *University Paris-13, France

268 Ly Thuong Kiet, Dist 10, Hochiminh city Vietnam, Tel: +84 903 787 989

Email: [email protected], [email protected], [email protected]

ABSTRACT : In the paper, an efficient super-resolution

image processing is presented. We have developed a fast

interpolation algorithm which is successfully

implemented on the ARM-9 microcontroller board,

AT91SAM9RL. The results have been compared to the

previous works on both with the Matlab-based

simulations and the hardware of DSP TMS320C5515.

The developed algorithm has shown up to be an efficient

method with less image-processing times for processing

large size color images in acceptable processing time.

The possible applications of super-resolution image

processing are to digital cameras or photos where high

resolution images are required for feature extractions.

KEYWORDS: Super-resolution image processing, DSP

board TMS320C5515, ARM-9 board AT91SAM9RL,

image processing times.

1. Introduction.

In the aim to produce supper-resolution (SR) images from

the given low-resolution (LR) images, there are some

works which have been done and achieved some certain

results [1,2,3,4]. Particularly for implementing the tasks

on hardwares, our DSP-research groups have successfully

implemented the super-resolution image processing on the

DSP kit TMS320C5515 of Texas Instruments since last

year [4], however a drawback of the work is that it

contains much of complex numeral operations. This

drawback is a comprehensive matter for improving the

algorithm in order to save the processing time whereas the

quality of the high-resolution (HR) images is still

remained. For this task, an efficient algorithm has been

developed for the image processing based on the rounding

of neighborhood pixels in the interpolation step of the

super-resolution image processing. The results shown in

both Matlab-based simulations as well as the hardware of

ARM-9 have successfully presented the efficient work for

the image signal processing. The improved algorithm

has been developed for low-resolution images (LR) to

achieve appropriated high-resolution images (HR) suited

to the tasks of SR image processing approach.

2. Supper-Resolution Process

Basically, there are three steps (Figure-1) to produce a HR

image [1,2,3,4] from a numbers of low-resolution (LR)

images as follows,

 Registration: the shift of the LR images in

horizontal and vertical dimension, the angle of

its rotation, and grid of pixels of LR images into

a same HR grid are estimated.

 Reconstruction: in this process, the

interpolation algorithm is used to interpolate the

remaining pixels of the HR grid

 Filtering: some image filters such as the median

filter and the Gaussian filter are applied to

smooth the HR images suffered from the

discontinuous reconstruction process.

Figure 1: Three steps [1] of producing a SR image:

Registration, Reconstruction, Filtering (De-blurring).

2.1 Registration

In [2], and [3], the authors have given details about the

registration process. For example, in [2] the shift

estimation is calculated in the frequency domain with the

change in the phase of the Fourier transform of LR

images while the rotation estimation is calculated by the

correlation between the “frequency contend H” of the

transform. [3], on the other hand, calculate the shift and

rotation change by solving a system of equations using

Taylor expansions and the iterations process. The

following approach is based on the frequency domain to

calculate the rotation and shift parameters of the

processed images.

Assume that )( 0 xf is the original image and )( 1 xf is the

shifted and rotated version of that original image [2, 4].

Then: ))(()( 101  xxRfxf (1)

Where:















 

v

h

x

x

x , 

















v

h

x

x

x

,1

,1

1 , 













  

11

11

cossin

sincos





R

.

Transform the equation (1) into the frequency domain

using the two-dimension Fourier transform: