An Adaptive Image-stitching Algorithm for an Underwater Moitoring System

An Adaptive Image-stitching Algorithm for an Underwater Monitoring System Regular Paper

Hengyu Li1,*, Jun Luo1, Chaojiong Huang1, Yi Yang1 and Shaorong Xie1

1 School of Mechatronic Engineering and Automation, Shanghai University, Shanghai, China * Corresponding author E-mail: [email protected]

Received 17 Jan 2014; Accepted 14 Aug 2014 DOI: 10.5772/58988 © 2014 The Author(s). Licensee InTech. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/3.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract Against the narrow range of sonar images for underwater monitoring, an adaptive sonar image-stitching algorithm is proposed in this paper. Compared with conventional sonar image-stitching algorithms, this algorithm can be applied to stitch sonar images with fewer features and better results. This paper first introduces the hardware platform of the underwater monitoring system and a 3D model of the underwater rotational mechanism. Next, related image pre-processing is conducted to improve its accuracy. A SURF algorithm is then applied to extract the speeded-up robust feature (SURF) points. Compared with a threshold, if the number of SURF points is larger, the SURF algorithm is applied to stitch the sonar images or else the phase correlation method is selected to mosaic the sonar images. Finally, a weighted smoothing fusion algorithm based on a maintained boundary is proposed to fuse the sonar images. The algorithm is verified to give good performance in mosaicing sonar images by an experimental study.

Keywords Underwater Monitoring System, SURF, Phase Correlation Method, Fusion Algorithm

1. Introduction

Currently the majority of countries are devoted to focusing on underwater monitoring. However, the acquisition of underwater images is performed in noisy environments with low visibility because natural light is not available - and even if artificial light is applied, the visible range is limited [1]. Acoustic cameras can provide extremely high resolutions (for sonar) and rapid refresh rates [2]. Therefore, sonar systems are widely employed to obtain images of the seabed or other underwater objects.

It is not conducive to monitoring objects or environments underwater through sonar equipment alone because the observation angle of high-resolution sonar equipment is so small that only part of the underwater scene can be observed, i.e., the horizontal view angle of DIDSON (dual-frequency identification sonar) is 28.8°. Therefore, most sonar equipment is often installed in rotational mechanism to obtain different scene information; as such, sonar images are stitched to expand the monitoring horizon of the underwater environment [3]. Thus, image mosaicing technology plays a significant role in underwater monitoring systems.

1Hengyu Li, Jun Luo, Chaojiong Huang, Yi Yang and Shaorong Xie: An Adaptive Image-stitching Algorithm for an Underwater Monitoring System

ARTICLE

Int J Adv Robot Syst, 2014, 11:166 | doi: 10.5772/58988

International Journal of Advanced Robotic Systems

There has been much research into optical image-stitching; however, little research has been devoted to stitching for sonar images. Image-stitching on an optical image can be divided into algorithms based on the pixel-level, the frequency power spectrum and features [4]. The calculation of the pixel-level algorithm is simple and fast, but the mosaicing result is not ideal if the images are particularly noisy or if there are not enough features. The calculation of the frequency power spectrum algorithm is also fast and noise correlation interference can be overcome, but it requires a sufficient overlap width. Meanwhile, feature points can easily be extracted by the feature algorithm, which is suitable for stitching the images with enough features but the cost of a heavy calculation load.

It is not appropriate to adopt a pixel-level algorithm for stitching sonar images with a lot of noise. In previous studies, many researchers have concentrated on the frequency power spectrum algorithm or the feature algorithm independently to mosaic sonar image. In the view of current study, the phase correlation method [5-7] is the most popular algorithm in the field of the frequency power spectrum algorithm while the SIFT (scale invariant feature transform) [8], FAST (features from accelerated segment test) [9] and ORB (oriented FAST and rotated BRIEF) [10] and SURF [11-13] algorithms enjoy tremendous popularity in the field of feature algorithms.

Several modified phase correlation algorithms based on the Fourier transform have been proposed in the literature [5-7]. The algorithms enable the measurement of translation, rotation and scaling factors between two images. The stitching resulting from this algorithm is good on paper. However, the precision of the stitching result for sonar images with sufficient features is lower than for the feature algorithm. The SIFT algorithm has been studied for extracting distinctive invariant features from images which can be used to perform reliable matching between different views of an object or scene [8]. Because the SIFT feature points are described as 128-dimensional vectors, the speed of this algorithm is not fast. Mair et al. [9] have described a novel corner detection approach, called the FAST (features from accelerated segment test) algorithm. The FAST algorithm only uses the surrounding pixels information to get the feature points, which is simple and relatively quick. However, the accuracy of the stitching result from this algorithm is worse than for the SIFT algorithm. The literature [10] has also described the ORB algorithm. An

ORB descriptor is obtained by adding rotational invariance to a BRIEF (binary robust independent elementary features) descriptor. This algorithm is described as a binary bit string and ach bit is defined by comparing the strength of point pair - the matching speed is greatly improved in this way. However, it is not suitable for stitching sonar images with few features. Another popular algorithm is the SURF algorithm [11-13]. SURF points are described as 64-dimensional vectors. Compared with the SIFT algorithm, this algorithm is faster. However, the stitching result of the SURF algorithm for sonar images without features is not good.

The underwater monitoring system described in the present paper is applied to monitor Huangpu River in Shanghai for security, which is more sandy, muddy and narrow than other aquatorium, such as an ocean or a reservoir. It is also frequently interfered with by tides, vessels and stormy waves, incurring low visibility and bad reverberation. Instances involving a great deal of background noise and fewer targets in the sonar image reflect the difficulty of obtaining a qualified mosaicing image. Moreover, the sonar device mounted on the USV is inevitably affected by waves during the operation process, resulting in the obtaining of fuzzy sonar images. Difficulties that arise might include large deformations after calibration or the absence of overlapping regions between images, ultimately leading to a failure or serious distortion when mosaicing.

Given the non-ideal result of separately applying an algorithm to stitch sonar images, this paper takes advantage of both the frequency power spectrum algorithm and the feature algorithm to stitch sonar images, and thus proposes an adaptive algorithm based on phase correlation and SURF for sonar image-stitching whose basic flow diagram is illustrated as in Figure 1. The adaptive algorithm proposed is an improved combination of the SURF algorithm and the phase correlation algorithm, reserving their respective remits. On the one hand, the feature points out of the detection zone (the fan zone) of sonar images are removed and inner feature points are double-checked by the RANSAC (Random Sample Consensus) algorithm. On the other hand, the translation parameter is obtained by computing the parameters of scale and rotation using log-polar transformation so as to improve the phase correlation algorithm.

Figure 1. The basic flow of the adaptive sonar image-stitching algorithm

2 Int J Adv Robot Syst, 2014, 11:166 | doi: 10.5772/58988

The structure of the paper is as follows. Firstly, a brief review of the hardware platform is given in Section 2. Preliminary tasks such as image pre-processing and feature point extraction are conducted in Section 3. Next, an adaptive sonar image-stitching algorithm for underwater monitoring system is proposed. The algorithm is divided into two parts, including the SURF and phase correlation algorithms, which are - respectively - described in Sections 4 and 5. Afterwards, the weighted smoothing fusion algorithm based on the maintained boundary is proposed to fuse sonar images in Section 6. In Section 7, we will present experimental results that confirm the effectiveness of the proposed method. Finally, the conclusions of the paper are stated in Section 8.

2. Hardware Platform Design

2.1 System Introduction

The underwater monitoring system is established based on a common boat, as shown in Figure 2. The system can be divided into three parts, comprising the fore, the cabin and the aft, according to the position of the boat. The fore, is equipped with sonar which has the capacity to detect targets accurately in muddy water and can be used for a wide range of fast patrols. System control and monitoring display are conducted in the cabin and the aft is equipped with an underwater robot that can move freely underwater. If the system finds a suspicious spot, the control system will control the underwater robot to conduct a close and thorough search.

2.2 3D Rotational Mechanism Design

Considering the cruising speed of the boat and the monitoring requirements underwater, the rotational stage should satisfy the following requirements, namely it should: (1) be able to rotate around three directions; (2) be water sealing, and (3) provide feedback information. As no commercially available rotational mechanism can meet these requirements, this paper presents a 3D rotational mechanism to extend the scope for monitoring, as shown in Figure 3.

Figure 2. Underwater monitoring system

Figure 3. 3D rotational stage

Figure 3 shows a 3D rotational mechanism, consisting of a yaw mechanism, a pitch mechanism and a roll mechanism, which can implement yaw, pitch and roll movements around three directions, driven by servo motors. Owing to requirement (2), several sealing devices are installed in mounting places. The 3D rational mechanism mounts attitude sensors to detect angle information, to meet the requirement (3). The designed 3D rational mechanism is capable of a maximum rotational speed of 60°/s and the adequate rotational ranges of yaw, pitch and roll range from -90° to 90°, from -70° to 10° and from -10° to 10°, respectively.

3. Image Pre-processing

Compared with optical images, sonar images have several characteristics [14]: (1) sonar images generally have a low resolution, vague details and poor identification because of the difference in size and the effect from the beam-forming arrays during image restoration; (2) Sonar images have strong Gaussian noise and a low SNR (signal-to-noise ratio) because of the absorption effect, the scattering effect and the convolution effect of light underwater.

In view of characteristic (2), the Gaussian smoothing algorithm is applied to smooth sonar images. Owing to characteristic (1), the algorithm based on stretching the grey-value of images [15] is adopted to improve image contrast and to enhance the image feature information of the target in monitoring area.

4. Feature Point Pair Calculation

4.1 Feature Points Extraction

This paper introduces the SURF algorithm [11-13] to extract feature points. The specific implementation steps of the algorithm are as follows:

(1) Convert the sonar image into an integral image. Anintegral image of an arbitrary point expressed as

),( yxXX = is represented as:


≤

=

≤

=Σ =

xi

i

yj

j

jiIXI0 0

),()( (1)

(2) Extract the interest points of the image by a Hessian matrix. A Hessian matrix of an arbitrary point in a scale σ is defined as:

=

),(),(),(),(

),( σσσσ

σ XLXLXLXL

XHyyxy

xyxx (2)

where ),( σXLxx represents the convolution of Gaussian second-order partial derivatives in direction x, and

),( σXLxy and ),( σXL yy have a similar meaning.

(3) Build a scale space. The algorithm applies increasingly large box filter templates to create an image pyramid.

(4) Determine the feature points. The algorithm defines local maxima as feature points by comparing a filter point with 26 points, including eight pixels in the same scale, nine points adjacent to an upper scale and nine points adjacent to a lower scale.

The feature point extraction of the SURF algorithm is improved by treating feature points out of a detection zone (a fan zone) as wrong points. The fan zone has a 28.8° fan angle and a radius which is the same as the image height.

4.2 Solving the Feature Vectors

Feature point gradients are first obtained in directions x and y as two axes in a new coordinate system. Next, the response to the feature points is mapped onto this coordinate system. The maximum of the accumulating wavelet responses in a certain range of direction is determined as the main direction of the gradient, whereby the feature points will be described in the main direction after that. The feature points are finally described as 64-dimensional vectors.

4.3 Matching the Feature Points

The Euclidean distance between the feature vectors of images is regarded as a similarity measurement of feature points. The algorithm first takes a feature point from a reference image and finds the nearest point and the second nearest point in the other image. The reference point and the nearest feature point are defined as a matching point pair if the proportional between the nearest distance and the second nearest distance is less than 75% [16]. All the feature points are traversed in the reference image and potential matching point pairs are found by this algorithm.

To improve the accuracy of the matching based on traditional feature point matching, feature point pairs need duplicate check because the resolution of the sonar

image is relatively low. Mismatched point pairs are rejected by RANSAC algorithm after coarse matching [17].

The prerequisite of using the SURF algorithm is that the number of feature points must be equal or else greater than four. If the number of feature points is greater than four, the SURF algorithm is applied to stitch the sonar images. If the number of feature points is less than four, the phase correlation method is introduced to stitch sonar images.

5. Image Registration Algorithm

5.1 Homography Matrix Solving by the SURF Algorithm

The sonar device nearly confirms the perspective transformation model during actual imaging. It should be noted that the homography matrix calculation confirms the condition whereby the number of feature points should be more than four. The relationship between two images can be concluded as a planar perspective transformation matrix, which is shown as follows:

21 HXX = (3)

where 1X is expressed as ( )Tyx 1,, 11 , 2X is expressed as ( )Tyx 1,, 22 , and ( )11, yx and ( )22 , yx are a matched point pair.

5.2 Calculating Parameters by the Phase Correlation Algorithm

5.2.1 Solving the Translation Parameters

The phase correlation method [5-7] is originally applied to register images that merely have a translation transformation.

The relationship of two image signals is assumed to be:

),(),( 0012 yyxxfyxf −−= (4)

where ox represents translation in direction x and oy represents translation in direction y . Thus, (4) is then

converted into (5) according to the Fourier transformation:

)(12 ),(),( oo vyuxjevuFvuF +−= (5)

where ),(1 vuF and ),(2 vuF respectively represent Fourier transformations of ),(1 yxf and ),(2 yxf .

The cross-power spectrum between ),(1 vuF and ),(2 vuF

can be calculated as follows:

)(

|),(),(|),(),(),(

2*

1

2*

1 ovyouxj

evuFvuFvuFvuFvuQO

+−

== (6)

where ),(*1 vuF is the complex conjugate of ),(1 vuF .


The inverse Fourier transformation of )( ovyouxj

e+−

is ),( oo yyxx −−δ , which is a two-dimensional pulse

function. Finally, the peak coordinate of ),( oo yyxx −−δ is determined as a translation parameter. The basic flow of the algorithm is illustrated as Figure 4.

Figure 4. The basic flow of the phase correlation algorithm

5.2.2 Solving the Scale and Rotation Parameters

The phase correlation algorithm can also be applied to calculate scale and rotation parameters [18]. However, the algorithm is applied to the logarithm polar form of the amplitude spectrum instead of directly to the image. The scale and rotation parameters of the image are converted to translation parameters by the transformation that will be presented below. The assumable relationship of two image signals is satisfied as:

)])cossin(,)sincos([),( 12

o

o

yyxxyxfyxf

−+−−+=

αασαασ (7)

where σ represents the scale parameter and αrepresents the rotation parameter. (7) can be converted to (8) according to the Fourier transformation:

|)]cossin(),sincos([),(

1

11

2),(2

2

αασαασσφ

vuvuFevuF vuj F

+−+

=−

−−−

(8)

where ),(2

vuFφ is the phase spectrum of ),(2 yxF .

The amplitude spectrum relationship is obtained from (8) and the result is illustrated as (9):

|)]cossin(),sincos([||),(|

1

11

22

αασαασσ

vuvuFvuF

+−

+=−

−− (9)

where |),(| 2 vuF is the amplitude spectrum of ),(2 yxf ,|),(| 1 vuF is the amplitude spectrum of ),(1 yxf . The

logarithmic polar coordinate transformation is then conducted for (9) and the result can be obtained as (10):

),(),( 12

2 κλαθσλθ −−= −plpl ff (10)

where )log(ρκ = .

The scale and rotation parameters between two images can be converted to translation parameters after the above transformation and then the phase correlation algorithm can be introduced to calculate translation parameters.

5.2.3 Solving the Translation, Scale and Rotation Parameters

As presented in 5.2.1 and 5.2.2, the translation parameter and the parameters of scale and rotation are calculated by the phase correlation algorithm and the transformed phase correlation algorithm, respectively. On this basis, we propose an improved transformed phase correlation algorithm to register sonar images. Specifically, the parameters of scale and rotation are calculated according to 5.2.2. Related image geometric transformations have been done to get only the translation relationship between the original image and the treated image. Next, the translation parameter is solved according to 5.2.1. The basic flow of the phase correlation algorithm is shown as Figure 5.

Figure 5. The basic flow of the phase correlation algorithm

6. Image Fusion

After the above processing, it is necessary to conduct coordinate mapping in the next step. To avoid black holes, inverse mapping is adopted in this paper. In addition, coordinate mapping will lead to an inevitable situation in which the integral points will be non-integer points after coordinate mapping. The double linear interpolation algorithm [19] is introduced to avoid this issue.

If images are fused directly by simple superposition, the spliced position will produce a conspicuous joint and blur overlap area. The weighted smoothing algorithm [20] is introduced to improve it. The idea behind the algorithm based on weighted smoothing is shown in Figure 6. The overlap of adjacent images 1f and 2f is in the interval

],[ 21 xx . )(1 xW and )(2 xW are weighted functions whose relationship is expressed as:

WixWxW /1)(1)( 12 −=−= (11)

where Wi ≤≤0 , W is the width of the overlapping portion.

Figure 6. The algorithm based on weighted smoothing


Finally, the pixel value of the fusion image f is calculated as:

)(),()(),(),( 2211 xWyxfxWyxfyxf += (12)

However, the above algorithm will inevitably cause the boundary of the overlapping area to become blurred. To enhance the integration of the image-stitching, the fusion algorithm based on weighted smoothing is improved and the author proposes a weighted smoothing fusion algorithm based on a maintained boundary. The basic flow of this algorithm is depicted as Figure 7.

Figure 7. The basic flow of the weighted smoothing fusion algorithm based on the maintained boundary

The specific implementation steps of the algorithm are as follows: (1) detect the edges of compass operator; (2) maintain the pixels if they belong to the boundary or else skip to the next step; (3) calculate the values of the pixels in the overlap portion according to (12).

7. Experimental Results and Analysis

7.1 Experimental Results of the Algorithm

The paper applies the proposed algorithm to stitch sonar images. A suitable threshold is obtained by continuous testing, which turns out to be 40 in our experiment. Two original images are shown as Figure 8(a) and (b), and the result of image-stitching is shown as Figure 8(c). The SURF feature point numbers of the two original images are, respectively, 43 and 42, which are larger than the threshold of 40. As such, the SURF algorithm is selected to mosaic the two sonar images.

Given that the feature point numbers of the two original images are larger than the threshold, this indicates that the images contain enough feature information. It is easy to see that the original images have a common characteristic object from Figure 8(a) and (b). As shown in Figure 8(c), there are eight feature point pairs mainly nearby the feature object, which are matched accurately. In Figure 8(d), it can be seen that the view field is widely expanded compared with the original images.

In addition, this algorithm is also applied to stitching sonar images which lack features. Two original images are shown as Figure 9(a) and (b). The numbers of the SURF points in the two original images are, respectively, 13 and 5, both of which are less than the threshold 40. As such, the phase correlation algorithm is selected to mosaic the two sonar images.

(a) (b) (c) (d) Figure 8. (a) and (b) The original image. (c) The result of image-geometric transformations of scale and rotation. (d) Stitched image.

(a) (b) (c) (d) Figure 9. (a) and (b) The original image. (c) The result of image-geometric transformations of scale and rotation. (d) Stitched image.


The relationship between original

ImagesSIFT algorithm

Phase correlation algorithm

The proposed algorithm

The images between Figure

9 (a) and (b)

Translation parameters: ( 5, 7) Scale and rotation:

1 and 0°

The images between Figure

10 (a) and (b)

Translation parameters: (30,

60) Translation parameters: (40, 3)

Scale and rotation: 1 and 0°

Scale and rotation: 1 and 10.8°

Table 1. The relationship between the original images calculated by different algorithms

As the numbers of feature points in the two original images are less than the threshold, this indicates that the images contain insufficient feature information. As seen in Figure 9(a) and (b), the two images have a common link. Figure 9(c) is the stitched result following the geometric transformations of scale and rotation from Figure 9(b) to (a). It can be concluded that the link is substantially coincident in the mosaic image and that the view field is widely expanded compared with the original images.

7.2 The Analysis of the Algorithm

In order to illustrate the effect of the algorithm more convincingly, this paper specifically carried out a set of comparative experiments. These comparative algorithms were selected from the algorithms based on the feature and transform domain, which enjoy significant prevalence in the image-stitching field. The algorithms based on SIFT and correlation are, respectively, adopted in the above two fields. This paper applies these algorithms to stitch the above sonar images in order to compare the effects occurring between the proposed algorithm and the classic algorithms. The relationship between the original images can be calculated by these algorithms, and the result is shown in Table 1.

The rotational movement of the sonar equipment can be measured by angle sensors. The sonar equipment rotates by around 3° from Figure 8(a) to (b) and the angle difference between Figure 9(a) and (b) is around 12°. In Table 1, the results of rotation calculated by the algorithm based on correlation are both 0°, far from the measured angles. Therefore, this algorithm is not appropriate for sonar image-stitching.

The case from Figure 8(a) to (b) has a scale of 1 and a rotation of 3°, with an approximate homography of:

0

0

0.99863 0.05233600.0523360 0.99863

0 0 1

xy

−

Compared with the SURF algorithm, the results of scale and rotation from the proposed algorithm are more accurate.

Similarly, the case of Figure 9(a) to (b) has a scale of 1 and a rotation of 12°, with an approximate homography of:

0

0

0.97814 0.20791170.2079117 0.97814

0 0 1

xy

−

The scale is 1 and the rotation turns to 10.8° calculated by the proposed algorithm, with a homography of:

0

0

0.98229 0.18738130.1873813 0.98229

0 0 1

xy

−

As to homography, the one from the proposed algorithm is closer to the standard value than the SIFT algorithm.

However, we cannot determine which is better between the proposed algorithm and SIFT from Table 1. Naturally, we switch emphasis to observing stitched images. The stitched images are shown as Figures 10 and 11.

(a) (b) (c)Figure 10. (a) The image stitched by the algorithm based on SIFT. (b) The image stitched by the algorithm based on correlation. (c) The image stitched by the algorithm proposed in this paper.

Compared with Figure 10(c), the bottom of Figure 10(a) is not fused well, even exhibiting some distortion. Similarly, compared with Figure 11(c), the common features of Figure 11(a) do not coincide well. It is concluded that the


proposed algorithm is better than SIFT. Figure 10(b) and Figure 11(b) are the fused results of the images stitched by phase correlation, which is still worse than Figure 10(c) and Figure 11(c). According to the above analysis, the proposed algorithm brings better results on sonar image-stitching.

(a) (b) (c) Figure 11. (a) The image stitched by the algorithm based on SIFT. (b) The image stitched by the algorithm based on correlation. (c) The image stitched by the algorithm proposed in this paper.

8. Summary

This paper proposes an adaptive sonar image-stitching algorithm for an underwater monitoring system which is based on phase correlation and SURF. It turns out to be effective and practical after experiments. An alternative algorithm can be applied to stitch sonar images according to the number of feature points. In other words, the algorithm employs SURF for sonar images with enough feature points while it employs phase correlation for sonar images lacking feature points. The paper also proposes a weighted smoothing algorithm based on a maintained boundary for image fusion which can enhance the integration of image-stitching. This algorithm takes advantages of the SURF and phase correlation algorithms to improve the accuracy and timeliness of sonar image-stitching.

9. Acknowledgements

This project is supported by the Key Projects of the National Natural Science Foundation of China (No. 61233010) and the National Nature Science Foundation of China (No. 61305106). The authors also gratefully acknowledge the helpful comments and suggestions of the reviewers, which have improved the presentation.

10. References

[1] K. Kim, N. Neretti, and N. Intrator (2005) Mosaicing of acoustic camera images. In: Radar, Sonar and Navigation. 2005; pp. 263-270.

[2] E. Belcher, B. Matsuyama, and G. Trimble (2001) Object identification with acoustic lenses. In: OCEANS, 2001. MTS/IEEE Conference and Exhibition. 2001; pp. 6-11.

[3] S. Xie, Y. Xu, J. Chen, W. Jin, and J. Luo (2012) Calibration and mosaicing of forward-scan sonar DIDSON images. In: Intelligent Computation Technology and Automation (ICICTA). 2012; pp. 439-442.

[4] B. Zitova and J. Flusser (2003) Image registration methods: a survey. Image and Vision Computing. 21: 977-1000.

[5] M. Druckmüller (2009) Phase correlation method for the alignment of total solar eclipse images. The Astrophysical Journal. 706: 1605-1608.

[6] H. S. Stone, M. T. Orchard, E. C. Chang, and S. A. Martucci (2001). A fast direct Fourier-based algorithm for subpixel registration of images. IEEE Transactions on Geoscience and Remote Sensing. 39: 2235-2243.

[7] N. Hurtos, X. Cuf, Y. Petillot, and J. Salvi (2012) Fourier-based registrations for two-dimensional forward-looking sonar image mosaicing. IEEE/RSJ International Conference on In: Intelligent Robots and Systems; pp. 5298-5305.

[8] D. G. Lowe (2004) Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision. 60: 91-110.

[9] E. Mair, G. D. Hager, D. Burschka, M. Suppa, and G. Hirzinger (2010) Adaptive and generic corner detection based on the accelerated segment test. In: Computer Vision–ECCV. 2010; Springer; pp. 183-196.

[10] E. Rublee, V. Rabaud, K. Konolige, and G. Bradski (2011) ORB: an efficient alternative to SIFT or SURF. In: Computer Vision 2011 IEEE International Conference. 2011; pp. 2564-2571.

[11] H. Bay, T. Tuytelaars, and L. Van Gool (2006) Surf: Speeded up robust features. In: Computer Vision–ECCV 2006; Springer; 2006; pp. 404-417.

[12] H. Bay, A. Ess, T. Tuytelaars, and L. van Gool (2008) Speeded-up robust features (SURF). Computer Vision and Image Understanding. 110: 346-359.

[13] E. Yong (2011) Investigation of mosaicing techniques for forward looking sonar. Master's thesis, Heriot-Watt University.

[14] M. Mignotte, C. Collet, P. Perez, and P. Bouthemy (2000) Sonar image segmentation using an unsupervised hierarchical MRF model. Image Processing. 9: 1216-1231.

[15] T. Arici, S. Dikbas, and Y. Altunbasak (2009) A histogram modification framework and its application for image contrast enhancement. IEEE Transactions on Image Processing; 18: 1921-1935.

[16] I. K. Jung, and S. Lacroix (2001) A robust interest points matching algorithm. Eighth IEEE International Conference on Computer Vision. 2: 538-543.

[17] M. A. Fischler and R. C. Bolles (1981) Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24: 381-395.


[18] Q.-s. Chen, M. Defrise, and F. Deconinck (1994) Symmetric phase-only matched filtering of Fourier-Mellin transforms for image registration and recognition. Pattern Analysis and Machine Intelligence. 16: 1156-1168.

[19] E. Sanville, S. D. Kenny, R. Smith, and G. Henkelman (2007) Improved grid-based algorithm for Bader charge allocation. Journal of Computational Chemistry. 28: 899-908.

[20] Z. Wang, D. Ziou, C. Armenakis, D. Li, and Q. Li (2005) A comparative analysis of image fusion methods. IEEE Transactions on Geoscience and Remote Sensing. 43: 1391-1402.


An Adaptive Image-stitching Algorithm for an Underwater Moitoring System

Documents

Transcript of An Adaptive Image-stitching Algorithm for an Underwater Moitoring System