chap1 thesis sample

6
Introduction Low resolution image optimization is an approach to enhance an image and produce an output which is of higher quality and better overall appearance. This optimization is highly invaluable and beneficial to law enforcement agencies and security institutions. In image processing, there are several approaches and techniques that can be utilized. These consist of image processing by enhancement, image processing by reconstruction and image processing by compression. Image processing by enhancement is subdivided into two approaches, Spatial filtering method and Contrast nhancement method. Spatial filtering method consists of a neighborhood and a predefined operation that is performed on the image pi!els encompassed by the neighborhood "#$. Contrast enhancements improve the perceptibility of ob% ect s in the sce ne by enha nci ng the diff ere nce in bri ght nes s bet wee n ob% ect s and the ir  bac&grounds "#$. Image processing by reconstruction minimizes the effect of degradations by filtering the observed image"'$. The effectiveness of image restoration is dependent on the filter design as well as on the e!tent and accuracy of the &nowledge of degradation process "'$. Image  processing by compression involves discrete cosine transform ()CT* that converts data into sets of frequencies and reducing the number of bits required and to represent an image "+$. There have been a numerous high grade image processing software that are being used in different fields of science researches. owever, each image processing has their limitations. The cubic spline "--$ is a general image interpolation function, but it suffers from lose of image details and blurring edges.  The recent attempts on developing on cubicspline interpolation are yet to be successful. Schreiber "-/$ proposed a sharpened 0aussian interpolation function to lessen information spillover among pi!els and enhance flatness in smooth areas . Schultz and

Upload: christian-arcelo

Post on 23-Feb-2018

227 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Chap1 thesis sample

7/24/2019 Chap1 thesis sample

http://slidepdf.com/reader/full/chap1-thesis-sample 1/6

Introduction

Low resolution image optimization is an approach to enhance an image and produce an

output which is of higher quality and better overall appearance. This optimization is highly

invaluable and beneficial to law enforcement agencies and security institutions.

In image processing, there are several approaches and techniques that can be utilized.

These consist of image processing by enhancement, image processing by reconstruction and

image processing by compression. Image processing by enhancement is subdivided into two

approaches, Spatial filtering method and Contrast nhancement method. Spatial filtering method

consists of a neighborhood and a predefined operation that is performed on the image pi!els

encompassed by the neighborhood "#$. Contrast enhancements improve the perceptibility of 

ob%ects in the scene by enhancing the difference in brightness between ob%ects and their 

 bac&grounds "#$. Image processing by reconstruction minimizes the effect of degradations by

filtering the observed image"'$. The effectiveness of image restoration is dependent on the filter 

design as well as on the e!tent and accuracy of the &nowledge of degradation process "'$. Image

 processing by compression involves discrete cosine transform ()CT* that converts data into sets

of frequencies and reducing the number of bits required and to represent an image "+$.

There have been a numerous high grade image processing software that are being used in

different fields of science researches. owever, each image processing has their limitations. The

cubic spline "--$ is a general image interpolation function, but it suffers from lose of image

details and blurring edges. The recent attempts on developing on cubicspline interpolation are

yet to be successful. Schreiber "-/$ proposed a sharpened 0aussian interpolation function to

lessen information spillover among pi!els and enhance flatness in smooth areas. Schultz and

Page 2: Chap1 thesis sample

7/24/2019 Chap1 thesis sample

http://slidepdf.com/reader/full/chap1-thesis-sample 2/6

Stevenson "-1$ put 2ayesian method to use for superresolution but hypothesized the prior 

 probability. Liu et al. "-#$ proposed a more structured retargeting method that performs a non

linear warp that accentuate interesting image features.

3esearch of 4ui 5ia and Shaogang 0ong focused on learningbased superresolution.

6hen employed to the human face, this is also widely &nown as 7hallucination8 "-$. Capel and

9isserman "-+$ made use of igenface which is chosen from training face database as model

 prior to constrain and superresolve lowresolution face images. :asztor "-'$ approached

learningbased superresolution differently. Learning from several general training images, they

tried to recover the lost highfrequency information from lowlevel image primitives. 2a&er and

4anade "'$ proposed a similar method of that with Capel and 9isserman, where they established

the prior based from a set of pi!el of training face images by pi!el using 0aussian, Laplacian and

feature pyramids. Liu and Shum ";-$ combined the image primitive technique of <reeman and

:C= modelbased approach.

In this thesis research, we will try to incorporate the techniques of image noise reduction

and learning based super resolution and add another illumination algorithm to even out the

radiance. These three different methods will be used together to come up with an optimization

technique. 6e intend to use <uzzy <ilter for noise reduction, Illumination Correction based on

Tone >apping for image illumination and lastly allucination by super resolution algorithm for 

recovering of lost highfrequency information occurring during the image formation process.

Page 3: Chap1 thesis sample

7/24/2019 Chap1 thesis sample

http://slidepdf.com/reader/full/chap1-thesis-sample 3/6

Background of the Study

= lot of researches and techniques in image processing have already been developed.

>ost of these studies came from the 5et :ropulsion Laboratory, >assachusetts Institute of 

Technology, 2ell Laboratories, ?niversity of >aryland and other research facilities "-$. ach

study proposes different types of algorithms and approaches.

@ne of these studies was from 6illiam T. <reeman, Thouis 3. 5ones and gon 0, :aztor.

=ccording to them, there are many possible ways to ma&e an image clearer and more distinct.

These can be done by sharpening and amplifying e!isting image details, gathering from multiple

frames and e!tracting a single highresolution frame from a sequence of lowresolution video

images. It all depends on what formula you will use in the process.

=nother study was from =ndrea :olesel, 0iovanni 3amponi and A. 5ohn >atthews";$. In

;BBB, they proposed a new method in image enhancement using =daptive ?nsharp >as&ing.

Their proposed approach uses an adaptive filter that controls the contribution of the sharpening

 path in such a way that contrast enhancement occurs in high detail areas and little or no image

sharpening occurs in smooth areas ";$.

In a separate study by :adilla, Sarte and uines about low resolution image

improvement, the results of the combination of the algorithmsD >ultiScale 3etine! =lgorithm,

0aussian <iltering with dge :reservation and <ace allucination Super 3esolution =lgorithm

has significantly improved the quality of face images ta&en from a low resolution camera.

0aussian <iltering with dge :reservation reduces the noise of an image and at the same time

 preserves the edges of the face which is a requirement for the ne!t algorithm, >ultiScale

3etine!. =fter 0aussian filtering has been applied, the difference in pi!el value of the image was

significantly reduced by ;E. This means that the edges have been emphasized. 3esults also

Page 4: Chap1 thesis sample

7/24/2019 Chap1 thesis sample

http://slidepdf.com/reader/full/chap1-thesis-sample 4/6

showed that >ultiScale 3etine! algorithm ma&es the illumination of an image even. ven

illumination provides constant pi!el value, which is very helpful in getting the average

 brightness of the image.

Problem Statement

The purpose of this study is to improve low quality images ta&en from standard low

resolution cameras. The sub%ects of these low quality images may be blurry or unidentified due

to the noise, low pi!el values, and uneven lighting of the image.

To address this problem, the proponents will combine three different algorithms namely,

<uzzy <ilter, Illumination Correction based on Tone >apping, and Super 3esolution =lgorithm,

which will be used to reduce the noise, correct the uneven illumination, and reconstruct the facial

features of the sub%ect of the low quality input image respectively. The output of this study is

e!pected to have reconstructed facial features and more recognizable sub%ect of the input image.

Significance of the Study

The output of this paper will significantly improve low resolution images into high

resolution images. 2y doing such, sub%ect of the images will be identified clearly and properly.

Thus, facial features and facial structures can be reconstructed which will ma&e the sub%ect of the

image clearer. This will be helpful to a lot of cases including but not limited to crime scene

investigation, surveillance cameras and other security related affairs.

:art of this study will improve e!isting algorithms since this paper will use different

algorithms such as <uzzy <iler, Illumination Correction based on Tone >apping and Super 

Page 5: Chap1 thesis sample

7/24/2019 Chap1 thesis sample

http://slidepdf.com/reader/full/chap1-thesis-sample 5/6

3esolution =lgorithm. These algorithms will be combined in order to test its capability to

improve the quality of the images. <urthermore, this study will also determine if the combination

of the used algorithms will enhance the quality of an image.

Scope and Limitation

 

Scope and limitations of the pro%ect are further e!plained at this part. >ain ob%ective of 

this research wor& is to improve image the quality of snapshot image ta&en from a low resolution

camera. The proponents ma&e this possible by introducing new combinations of algorithm.

  <irst, the proponents must gather low resolution input images. <or the resolution, the

 proponents used ;1B!/;B low resolution cameras for capturing input images. =ll gathered low

resolution images will undergo noise reduction process. =fter noise reduction process, image

illumination will be ad%usted for better image clarity.

<or gathering input data, sub%ect person must be - meter away from the camera. The proponents

will also gather -BB different high resolution photos together with its corresponding -BB low

resolution images of different persons. The camera that was used for ta&ing low resolution

 pictures was a Fo&ia /;-B with a camera resolution of ;GB!/;B at .- megapi!els. @n the other 

hand, the camera that was used for ta&ing high resolution pictures was an Samsung S1 with a

camera resolution of ;GG+!/;#G at + megapi!els. Hedit depending on the device to be usedH

<or the orientation, the proponents gathered images that are standing or sitting for 

 position and facing the camera, has center light for light source and people smiling or neutral for 

emotions.

 

Page 6: Chap1 thesis sample

7/24/2019 Chap1 thesis sample

http://slidepdf.com/reader/full/chap1-thesis-sample 6/6

0athered <ace images from the databases have different area of image occupied and sizes

 by face varies considerably.

 

<or limitations, the proposed thesis cannot process images ta&en during bad weather 

conditions, such as heavy rain and foggy environment. Images that have too much sunlight

e!posure cannot be processed. =nother limitation of the paper is sub%ect person who wears hat,

scarf, hood and any face garments cannot be reconstructed and processed. Sub%ect person must

 be directly facing the camera and only one person can be processed in a given time.