Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h, Vo lume 3, Issue 3, Marc h-2012 1

ISS N 2229-5518

A Review and Analysis of Image Misalignment

Problem in Remote Sensing

Bukenya Faiza, Siti Sophiayati Yuhaniz, Siti Zaiton Mohd Hashim, Kiweewa Kalema AbdulRahman

Abs tractIn change detection analysis, the accuracy of matching techniques depend solely on the accuracy of correction methods (such as geometric correction method, intensity variation methods) used bef ore the actual align ment is perf ormed. When the poor cor rection methods are used during image processing, errors such as matching errors, localisa tion error and alignment error, may arise during the detection and matching of image f eatures (such as points,contour) hence causing image misalign ment w hich may in turn aff ect t he change detection accuracy. This paper review s some of the challenges f aced during image registration together w ith the current image registration techniques. It also review s the errors associated w ith most of the image registration techniques due to use of poor or inappr opriate image processing techniques.

Inde x TermsImage Misalignment, Image Reg istration, Change Detection, Re mote Sensing.

1 INTRODUCTION

mage registration is the rationale of overlaying two or more images of the same scene captured at different times, from different viewpoints, and by different sensors or same sensor [12]. These images are often called sensed images or
target image and reference images or source image.
Majority of the authors reported that change detection ac-curacy depends solely on image registration accuracy. According to [1][6][11], it is so important to achieve accurate image registration results in order to avoid induction of false and missed changes that may lower down change detection accuracy.
However literature shows that improvements have been made but there is still much effort required to further improve on some the existing image registration techniques. This paper reviews some of the challenges faced during image registration together with the current image registration techniques. It also reviews the errors associated with most of the image registration techniques due to use of poor or inappropriate image processing techniques and suggest more efficient methods for alignment as future work. Section 2 describes some of the challenges faced during image registration. Some the current image registration tech -niques are also reviewed. Section 3 presents several errors that are faced during image registration. Section 4 concludes this paper.

2 IM AGE REGISTR ATION

2.1 The Challenges

This subsection describes and examines the causes of the challenges faced during image registration process. More emphasis is put on the causes of the problems. Some of the image registration problems include geometric distortion, differences in images such as intensity, contrast, resolution, sensor and environmental noise.
a) Geometric distortion
All remote sensing images regardless of the type of sensor used during image capturing, they are subject to geometric distortion. This is due to an attempt of accurately representing three dimensional images as two dimension images hence causing relief displacement. The two type’s geometric distor- tion includes:
i). Tangential scale distortion
This type of geometric distortion is caused by the rotati on of
the scanning optics. The problem occurs When the scanning mirror rotates at a constant speed, causing the IFOV of the sensor to move faster (relative to the ground) and closer to edges hence scanning a larger area as it moves closer to the edges .
ii). Skew distortion
The problem is seen to result in the instance of satellite orbit,
where the eastward rotation of the Earth causes the sweep of
scanning systems to cover an area slightly to the west of each previous scan hence resulting in to an image s kewed across the image.
b) Difference in images.
The difference in the images (such as resolution, contrast non
uniformities, illumination, and intensity) may occur due to being captured at different wave length by different sensors at different time. Hence making it difficult to compare images with different features such as image of different intensities [7], it’s not easy to construct descriptors [4] that can provide global information about feature point, in order to effectively differentiate signals in feature descriptions and also help dur- ing matching process [10].
Table 1 shows the challenges faced by most of the re-searchers during image registration. However these challenges are the major causes of image misalignment problem in change detec- tion analysis.

Table 1: Majors challenges in image registration

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3, Marc h-2012 2

ISSN 2229-5518


The next section gives an overview of the currently pro-posed image registration techniques to tackle some of the major problems that affect change detection accuracy.

2.2 Current Image Registration Technique s

Several image registration techniques have been proposed to deal with several challenges (such as geometric distortion, intensity variations, sensor noise and environmental noise) faced during image registration, but literature show s that image misalignment still exist with all the image registration techniques.
One of the most applied image registration
techniques proposed are feature based, they involve matching features such as points, edges during image registration.
Among the image registration techniques, active contour
segmentation and mutual information method produces better image registration results compared to all methods that use point feature this is because active contour segmentation involve matching edges rather th an point which saves time. It is capable of extracting geometric details of the feature edge compared to methods that use point feature to register image, also it is less sensitive to errors (localisation errors, matching errors, alignment errors),below a re some of the current image registration techniques.
1) A fully automatic and fast non-rigid image registration technique [10]
The technique was proposed to deal with difference in illuminations, difference in resolutions of the images. In this technique point features of the images under study are auto- matically detected by the wavelet pyramid and selected by the Harris corner method. Then later extracted point features of the two images (sensed image and reference image) are matched using scale invariant feature transforms. According to [10], the method is fast and robust although it is not capable of stating the properties of geometry in the SIFT (scale invariant Feature transform) key point location.
2) A wavelet multi-resolution based feature extra ction tech-
niques [2].
The technique was presented to increase relief displace-ments, increase the searching speed feature extraction method. In this method a wavelet multiresolution property is used to automatically detect the feature points in the two i mages under study. The extracted feature points are then matched using the normalised cross correlation. The method is fast but it only applies to grey scale feature points and also it does not resolve the geometric distortion but however [2] recommends the use of a true orthorectification technique to solve the geometric distortion in images under study.
3) A combination of active contour segmentation and mutual information [9]
Here the active contour segmentation method was pre-sented to improve the searching speed of the mutual informa -tion. The active contour segmentation method partitions the image in order to extract the edges of the images under study. Then later corresponding edges of the images are matched using mutual information method. The method is computationally fast. But not accurate enough due to error caused due to failure to deal with the geometric distortion.
4) Automated Inter-sensor or Inter-band Satellite Image Reg- istration System [7]
The technique was proposed to eliminate sensor and envi- ronmental noise, contrast non -uniformities, and inter-sensor and inter-band intensity mapping differences. In this tech - nique a novel modified German–McClure M-estimation scheme that uses a robust phase-adaptive complex wavelet feature representation to robotically match control point. Then an iterative refinement scheme is used to improve the control point pair localization. According to the author the technique produces better results as compared to state-of-the- art M-SSD and ARRSI registration algorithms [7].
5) Robust scale invariant feature transform descriptor [4]
It was originally proposed to deal with huge variation in both geometry and intensity in the images under study. The novel proximity matrix is used to extract the feature point from the images under study, and then later robust scale invariant feature transform descriptor is used match the feature points of the two images. The proximity matrix is used during feature extraction because the matrix combines the geometric information of feature points with the gradient information of feature points’ neighbourhood. According to [4], the method is computationally fast compared to the scale invariant feature transform descriptor although there it still faces a problem of large view point difference between images in a particular degree which leads to inaccuracy during matching process hence resulting into image misalignment.

2.3 Summary Of The Existing Image Registration

Techniques

Table 2 shows some of the current image registration tech-

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3, Marc h-2012 3

ISSN 2229-5518

niques that has been discussed previously, emphasizing the strength, weakness, type of remote sensed imagery used.

Table 2: Comparison of existing image registration tech- niques


One of the most applied image registration techniques pro- posed are feature based, they involve matching features such as points, edges during image registration. Among the image

———— ——— ——— ——— ———

Bukenya faiza,Faculy of Computer Science and Information sy s- tem,Universiti Teknologi Malaysia,

Siti Sophiayati Yuhaniz,Faculty of Computer Science and Infor- mation system,Universiti Teknologi Malaysia

registration techniques, a combination of active contour segmentation and mutual information method produce better
image registration results compared to all methods that use point feature this is because active contour segmenta tion involve matching edges rather than point which saves time. It is capable of extracting geometric details of the feature edge compared to methods that use point feature to register image, also it is less sensitive to errors (localisation errors,
Table 2 shows that the method that use active contour seg-mentation as feature extraction method produces better image registration results as it is fast compared to all methods that use point feature because active contour segmentation is good at extracting geometric details of the feature edge. Methods that used wavelet pyramid to extract are computationally fast, can deal with illumination variation in images but not good at extracting the geometric details of the feature points. Therefore a combination of active contour and wavelet pyramid would yield better image registration results.
The next section describes errors the limit the performance of
the image registration techniques.

3 TYP ES OF ERRORS IN IMAGE REGISTRATION TECHNIQUES

1) Alignment Error
This error occurs if the two images under study are not prop-erly positioned to each during image warping i.e. the sensed image is not properly positioned against the reference image during image warping. Image warping refers to man i- pulation of an image during image registration such during the correc-tion of geometry distortion.
Alignment errors are some of major problem faced during image registration due to using inappropriate mapping model for geometric distortion and poor calculation of the model parameters.
In order to estimate alignment accuracy, a consistent
check is used whereby two or more different image registra-
tion methods are applied on a pair of images under study. The result is achieved by applying the two methods then com- pared. If such results are similar or nearly similar, alignment is considered to be accurate otherwise the alignment is inacc u- rate.
2) Matching error
This error occurs in instances where the two images under
study are wrongly matched, that is where the control points of
image 1 are matched with wrong control poin ts of image 2. This is a common mistake made in most of image registration techniques. In order to identify mistakes, a consistent check can be used.
During this process, two different matching methods
are applied on the two images where control point of image 1
are matched with the control points of image 2. The result is from the two methods are compared such that if they have a matched pair in common, that indicates that pair is valid and therefore the remaining pairs are excluded for further processing by either using different methods or by applying a

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3, Marc h-2012 4

ISSN 2229-5518

cross validation to identify invalid control point pair.
In cross validation, the mapping parameters are ca l-
culated on the excluded pairs (considered invalid) in order to determine how the excluded points can be mapped to each other through the use of a model. Such that if the displac e- ment is below the threshold, then it’s considered to be a valid pair otherwise it’s invalid.
3) Localization Error
This error occurs when an image 1 is involved with wrong co-
ordinates in image 2. This type of error is common in cases
where the control points are detected using poor or inappro- priate detection algorithms.
This type of error cannot be measured directly on the
im-age as compared to other types of errors. It c an only be meas-ured by using ground truth data sets; ground truth is a process in which a pixel on a satellite change image is com- pared to corresponding pixel of the real image in order to veri- fy the contents of the pixel on the change image.
Localization error can be reduced through using an
appropriate feature detector and increase in localization errors on control points does not necessarily mean inaccuracy. Some- times it is needed in order to achieve accurate results. Table 3 summarizes errors that are rampant in most of image registra- tion techniques.

Table 3: Common errors faced during image registration


Table 3 shows the errors that are due to the use of poor or inappropriate correction methods (geometric distortion correction method, illumination variation correction method, intensity variation correction methods), this minimises the performance of the feature extraction and matching methods during image alignment process. Therefore better techniques
are required to solve the problems (such as geometric
distortion, varying illumination) before image alignment. And also better feature extraction and matching techniques are required for better image alignment process.

4 DISCUSSION AN D CONCLUSIONS

The performance of the extraction and matching techniques depend on the condition of the images under study, it would be better if thorough image processing is perfomed before i m- ages are subjected to alignment process. This is will improve on the extraction and matching process since images will be in better condition for a better image alignment.
Therefore much attention should be drawn on the image
processing part since it is where the image misalignment er- rors arise from. Since methods that adopt segmentation[9] during image registration are involved with less segmentation error and fast compared to other methods[4][7][10], better re- sults may be obtained if better image processing techniques is done. Thus, better segmentation process during image align- ment may produce image that is less sensitive to errors and noise. Therefore we suggest that both of these processes should be taken in to consideration for a better image align- ment, which is the focus of our future work.

5. REFERENCES

[1] Dai, X. L. and S. Khorram (1998). "The Effects Of Im- age Misregistration On The Accuracy Of Remotely Sensed Change Detection." IEEE Transactions on geoscience and re- mote sensing 36(5): 1566-1577.
[2] Hong, G. and Y. Zhang (2008). "Wavelet-Based Image Registration Technique For High -Resolution Remote Sensing Images." Computers and geosciences 34(12): 1708 -1720.
[3] Jay Gao.(2009),”Digital Analysis Of Remotely Sensed Imagery”, isbn: 9780071604659,division: professional,pub date: feb-09,pages: 674,edition: 01
[4] Liu, X., Z. Tian, (2011) "A Novel Adaptive Weights Proximity Matrix For Image Registration Based On R -Sift." AEU - international journal of electronics and communications (0).
[5] Marchesi, S., F. Bovolo, et al. (2010). "A Context-
Sensitive Technique Robust To Registration Noise For Change
Detection In Vhr Multispectral Images." image processing, IEEE Transactions on 19(7): 1877 -1889.
[6] Townshend, J. R. G., Justice, C.O., Gurney, C.,
Mcmanus, J., “The Impact Of
Misregistration On Change Detection,” IEEE Transactions on
geoscience and remote sensing, vol. 30, no. 5, 1992, pp. 1054 -
1060.
[7] Wong, A. and D. A. Clausi (2010). "AISIR: Automated
Inter-Sensor/Inter-Band Satellite Image Registration Using
Robust Complex Wavelet Feature Representations." Pattern recognition letters 31(10): 1160-1167.
[8] Yan Gao and Jean Francois Mas , 2008. “A Compari-
son Of The Performance Of Pixel Based And Object Based

IJSER © 2012

http :// www.ijser.org

Inte rnatio nal Jo urnal o f Sc ie ntific & Eng inee ring Re se arc h Vo lume 3, Issue 3, Marc h-2012 5

ISSN 2229-5518

Classificatsifications Over Images With Various Spatial Res o-
lutions.” Online journal of earth sciences, 2: 27 -35.
[9] Yang, Y. and X. Gao (2009). "Remote Sensing Image Registration Via Active Contour Model." AEU - international journal of electronics and communications 63(4): 227 -234.
[10] Yu, L., D. Zhang, et al. (2008). "A Fast And Fully Au- tomatic Registration Approach Based On Point Features For Multi-Source Remote-Sensing Images." computers & geosciences 34(7): 838-848.
[11] Yuhaniz, S. S. and T. Vladimirova (2009). "An On-
board Automatic Change Detection System For Disaster Mon i-
toring." international journal of remote sensing 30(23): 6121 -
6139.
[12] Yuhaniz, S., T. Vladimirova, et al. (2005). “Embedded
Intelligent Imaging On-Board Small Satellites.” advances in computer systems architecture, proceedings. srikanthan, j. xue and c. h. chang. berlin, springer-verlag berlin. 3740: 90-103.
[13] Bunting, P., F. Labrosse, et al. (2010). "A Multi - Resolution Area-Based Technique For Automatic Multi-Modal Image Registration." Image and Vision Computing 28(8): 1203 -
1219.

IJSER © 2012

http :// www.ijser.org