Firefly Algorithm for Optimized Non-rigid Demons Registration

1

Firefly Algorithm for Optimized Non-rigid Demons Registration

2

Sayan Chakraborty1, Nilanjan Dey2,*, Sourav Samanta3, Amira S. Ashour4, Valentina E.

3

Balas5

4 5 6

1

Department of CSE, Bengal College of Engineering and Tech., Durgapur, West Bengal, INDIA. Email:

[email protected]

7 8 9 10 11

2*

12

University, EGYPT. College of CIT, Taif University, KSA. Email: [email protected]

13 14

5

Department of Information Technology, Techno India College of Technology, Kolkata, INDIA. Email: [email protected] 3 Department of Computer Science & Engineering, University Institute of Technology, Burdwan, West Bengal, INDIA. Email: [email protected] 4 Department of Electronics and Electrical Communications Engineering, Faculty of Engineering, Tanta

Faculty of Engineering, Aurel Vlaicu University of Arad, ROMANIA. Email: [email protected]

Abstract: Videos have vital applications in numerous real-time aspects such as teaching, learning, communication, computer vision and medicine. Typically, video registration is entailed to describe a part of the scene/object in the video frame or to localize an object in the frame relative to a fixed reference system. Since, the semi-local transformation generated by the B-Splines registration was solved using demons algorithm. Thus, the current study is concerned with demons algorithm based image registration for a fully local transformation model. The demons registration is optimized using Firefly algorithm (FA) to optimize the velocity smoothing kernels of the demons registration considering the correlation coefficient as a fitness function. Afterwards, the proposed system performance using demons algorithm based FA is compared to the Particle Swarm Optimization (PSO). The experimental results proved that the proposed system based FA achieved correlation value of 0.6108 compared to demons registration with default parameters that provided 0.4468. Additionally, the FA based optimization framework was more stable and produced superior results than the PSO based optimization framework. Besides, the FA algorithm converged faster than the PSO. Keywords: Video registration, demons registration, non-rigid video, Firefly algorithm, fitness function, correlation coefficient, particle swarm optimization.

15 16 17

1. Introduction

18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33

Image processing is a wide domain that has various processes including enhancement, feature extraction, registration, segmentation, pattern matching, classification, fusion, morphology analysis and statistical measurements (Bulsara et al., 2011). Thus, researchers are interested to study miscellaneous image processing applications and optimization algorithms (Dey et al., 2012a; 2012b; Sharma, T. K. and Pant, 2013; Roy et al., 2014; Mathew et al., 2014; Jakšić et al., 2015; Pal et al., 2015; Nandi et al., 2015; Gospodinova et al., 2015). Typically, image registration is a significant process in many potential applications (Chowdhury et al., 2014; Araki et al., 2015). It is the process of aligning two images that correspond to the same scene into a common coordinate system to locate a definite object, by comparing pixel by pixel value of both the sensed and referenced images. These aligned images may be captured from different imaging devices or at different time instances. Since, the video contents are separated into multiple image frames. Thus, in video registration, each moving image (target image) is mapped with reference (fixed) image. Generally, the referenced image is kept fixed and is used as basis for the sensed image. Video sequence registration involves the spatial transformation recovery besides the temporal alignment recovery between the two videos. Features are extracted from every frame of the two video sequences. Afterward, the features from a frame in the first sequence are to be matched with the corresponding frame’s features in the second sequence. 1

Chakraborty et al.

1 2 3 4

Vigorously in the recent decades, non-rigid image registration techniques have been extended to develop the registration accuracy. Such methods are the similarity criterion (intensity-based, landmarks, surfaces) approaches, different deformable models (non-rigid models using splines, rigid/ affine models, wavelets, dense/non-parametric) and regularity constraints as well as adding different constraints on the transformation.

5 6 7 8 9 10 11 12 13

Typically, effective motion correction requires non-rigid image registration, which facilitates more flexible matching of the local details between two images than rigid registration. B-splines can fit the non-rigid deformations model, where a grid of fixed window size is formed to assist the mapping of the target image according to the reference image. Non-rigid image registration methods can be extensively grouped into two classes, namely intensity-based technique and feature-based technique (Crum et al., 2004). Feature-based techniques provide precise models without manually/ automatically select the control points/features. It utilizes the attribute vector model to match related features between images. Thus, the registration process can be considered feature matching problem. To match the images, the feature-based techniques depend on a relatively small number of feature points.

14 15 16 17 18 19 20 21 22 23 24 25 26

On the contrary, the intensity-based techniques deal directly with the image intensity and measure the intensity similarity, and then adjust the transformation till the similarity measure reaches its optimal value. Thirion (1998) presented the demons algorithm as a delegate to the intensity-based techniques and similar to an optical flow approach for small displacements. It is based on the standard of intensity conservation between images as it is derived from the optical flow model. In the demons algorithm, each pixel can have its own displacement. Conversely, in real images, accurate intensity matches do not essentially imply good registration of the underlying images. Therefore, the energy function being optimized in the intensity-based techniques suffers extensively from local minima. Demons is a representative form for the fluid registration algorithm that have different parameters, such as velocity field smoothing kernel and the alpha noise constant. Thus, it requires optimization algorithms to attain the parameter’s optimal values. In addition, in the video segmentation, the relationship between frames of the two video sequences is unknown. This relationship can be achieved using all potential selections of the frame pairs from the two video sequences. This broad search scheme will lead to the optimal result, but will suffer from increased computational cost to solve this brute force optimization.

27 28 29 30 31

Consequently, optimization algorithms should be employed to search for specific optimal solutions using an objective/multi-objective function. Meta-heuristic algorithms, such as the Particle Swarm Optimization (PSO), Cuckoo Search (CS) and Firefly Algorithm (FA) are high level approaches for exploring the search spaces. Thus, in the current work, the FA is employed to optimize the velocity smoothing kernels of demons registration.

32 33 34

The remaining sections are structured as follows. Section 2 included the related works followed by the material and methods in section 3. The proposed method and the discussed results are introduced in sections 4 and 5; respectively. Finally, the conclusion work is presented in section 7.

35

2. Related Works

36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51

The main goal of image/video registration is to estimate the transformation, which registers the reference image and the target image. This process can be performed by optimizing the metric function (similarity measure). The gradient decent scheme was used for medical image registrations (Brown, 1992), which considered local optimization technique. Since, the transformation parameters are generally non-convex and irregular. Thus, to avoid the local minimum, these schemes require good initial values for estimation. To overcome the problem of the local minimum, Rouet et al. (2000) employed the genetic algorithm (GA), which is a powerful global optimization technique, to optimize medical image registrations. However, the GA suffers from some disadvantages, such as consuming large computation time and lacking of fine tuning potentials. Consequently, Wachowiak et al. (2004) adapted the PSO algorithm for single-slice 3D-to-3D biomedical image registration. The experimental results of optimizing the normalized mutual information similarity metric were compared to other evolutionary strategies. The hybrid PSO method produced accurate registrations compared to the evolutionary strategies in terms of the convergence. The authors demonstrated that the PSO approach along with evolutionary algorithms and local methods were functional in image registration. Klein et al. (2007) compared the performance of eight optimization methods: gradient descent, nonlinear conjugate gradient, quasi-Newton, simultaneous perturbation, Robbins–Monro, Kiefer–Wolfowitz, and evolution strategy of medical image registration. The optimization techniques for registration were tested on manually deformed computed 2

1 2

tomography (CT) images of the heart, CT chest images and on Magnetic resonance (MR) images. The results demonstrated that the Robbins–Monro technique was the best in several applications.

3 4 5 6 7 8 9 10 11 12

In 2008, Chen et al. applied PSO to 2-D non-rigid image registration, which proved its effectiveness. In (Mohamed and Hamza, 2010) an optimized entropy based image registration algorithm has been proposed for medical image registration. The authors employed a modified instantaneous perturbation stochastic approximation algorithm in the optimization process. In (Meskine et al., 2010), a GA based rigid image registration has been suggested for satellite/radar images registration. Afterward in (Zheng and Tong, 2011), a hybrid PSO algorithm was applied on mutual information to optimize image registration. Singhai and Singhai (2012) presented a dominant search strategy based on the GA to register satellite images. The authors applied mutual information (MI) to calculate the statistical dependence of the information redundancy between the image intensities of matching voxels in both reference image and floating image. The results established that this conducted approach overcame the limitation of local maxima with the desired speed and accuracy.

13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31

Bejinariu et al. (2013) suggested a parallel method for the pixel intensity based image registration (IR) problem on multi-core processors. A geometric transform consisted of two classes of bio-inspired algorithms, namely Bacterial Foraging Optimization Algorithm (BFOA) and Genetic Algorithm (GA) were employed. The optimal transform was applied to a target (source) image in order to align it to a reference image by maximizing a similarity measure. The MI was used to evaluate the IR quality and the consumed processing time. Hwuang et al. (2013) proposed a novel approach based on discrete optimization that outperformed other approaches for deformable CT introduced the Anisotropic smoothing regularizer (AnSR). The edge-detection and de-noising within the Demons framework were conducted. The authors used AnSR within the demons algorithm and executed pair-wise registration on 2D synthetic brain MRI with and without noise after inducing a deformation. The results illustrated that by selecting the displacements in the deformation field, AnSR surpassed both GaSR and no regularizer (NoR) in terms of normalized sum of squared differences (NSSD). In (Caspi et al., 2006), feature trajectories have been engaged instead of features from every frame for video registration. The authors extracted the feature points from the first frame and tracked all over the video. To improve the spatial/ temporal transformation between the video sequences, the optimization problem has been solved. Merritt et al. (2013) depicted a CT-video registration scheme for image alignment and image-based rendering. The authors proposed an inverse-compositional framework using a gradient-based optimization procedure. The experimental tests proved the robustness and accuracy of the proposed system for both single frames and continuous video sequences. The benchmark timing tests specified that the conducted method can run continuously at 300 frames/sec.

32 33 34 35

The foremost survey included related work on image/ video registration problem for both rigid and non-rigid scenes, jointly with introduced previous studies on the optimization concept to support image registration. Consequently, it is obvious that no preceding studies have been employed the Firefly algorithm (FA) in the domain of demons registration optimization.

36 37

3. Material and Methods

38 39 40 41

The proposed framework includes three major parts: (i) Binning, (ii) Demons registration and (iii) Firefly algorithm (FA) to optimize the velocity smoothing kernels of the demons registration.

42 43 44 45 46 47 48 49 50

The proposed system defragmented the video content into multiple frames. Then, apply a binning strategy prior to the registration process in order to define the middle frame of the bin being the target frame. The binning selection strategy is executed to solve the problem created by spatio-temporal localization changes of the frames. This problem occurs during the frame registration process, where a correlation between two consecutive frames exists. By minimizing the distance between moving/ target images, registration process sensitivity is maximized. Since, any miss selection of the target frames can direct to insignificant registration. Thus, the current study proposed the use of a bin selection strategy for selecting the number of frames to be used as a set for registration after dividing the entire video m frames into a bin-size of n to solve this mentioned problem. The target frame

3.1 Binning

3

Chakraborty et al.

1 2 3 4

demotes either the middle (nth/2) or the common frame in both halves. This process carried on till all the bins are processed.

5 6 7 8 9

To resolve the deformable image registration, demons framework is employed as a classical optimization scheme. Through transformation, each pixel has a displacement vector, thus the demons framework is able to solve the semi-local transformation problem generated by B-Splines registration. The deviator displacement D between the moving image (t) and the reference (fixed image) (r) is computed using the following iterative formula:

3.2 Demons registration

Din, j  Din, j 1  10 11

(tin, j 1  ri 0, j ) ri 0, j |  ri 0, j |2  | (t in, j 1  ri 0, j ) |2 '

(1)

This iterative formula is subjected to the following initial condition:

12 D

0 i, j

 0,

t i0, j

 tˆi , j , ri 0, j  13 rˆi , j

(2)

14 15

Here, i, j = 1, 2, 3,….., N, while the original static and moving images intensities at the corresponding pixel are

16

3.3 Firefly Algorithm

17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35

Meta-heuristic algorithms, nature-inspired with multiple interacting agents, have vital role for contemporary global optimization algorithms, soft computing and computational intelligence. Among the new meta-heuristic, it has been exposed that FA is very efficient with multimodal and global/ local optimization problems. The FA is based objective function that can find an optimal solution to a problem by iteratively try to enhance a candidate solution considering a specified measure of solution quality. It was developed by Xin-She Yang (2008) based on the flashing patterns and behavior of fireflies (Yang and He, 2013). The three basic rules for the FA modern meta-heuristic algorithm are stated as follows based on the firelies characteristics:

36 37 38 39 40 41 42 43

Here, β0 refers to the maximum attractiveness at r = 0. γ denotes a fixed light absorption coefficient, which controls the decrease in the light intensity. In practice, the value of γ is established by the system’s characteristic length that to be optimized within the range of 0.1 to 10 from the full range of γ ∈ [0,∞]. Meanwhile, the distance over which the attractiveness changes significantly is known as the characteristic distance Г. For certain characteristic, the length scale Г in an optimization problem, while the parameter γ can initialize by the following value:

represented by tˆi , j and rˆi , j ; respectively, where N is any positive integer.

 

Fireflies are unisexual, as they move towards more brighter/ attractive fireflies regardless of their sex. The attractiveness is related to the brightness, while the brightness is inversely proportional to the distance among the fireflies. Thus, for any two flashing fireflies, the less bright firefly moves in the direction of the brighter one. If there are no brighter firefly than the particular one, the firefly will move randomly.  The objective function setting establishes the brightness of a firefly. Generally, based on the problem domain, the value of the objective function is proportional to the brightness. The significant concerns in the FA are: the attractiveness formulation and the light intensity variation. The firefly’s attractiveness is proportional to the light intensity that seen by nearby fireflies. The attractiveness function β(r) with the distance between two the adjacent fireflies r, is given by:  (r )   0 e   r

m

, (m  1)

 4

1 m

(3)

(4)

1

For fixed γ, the characteristic distance is:    1  1 when m  

2 3

Though the distance between any two fireflies i and j at xi and xj are expressed as:

4

rij  || xi  x j || 

d

 (x

i ,k

 x j ,k ) 2

(5)

k 1

5 6 7 8

Where, xi,k is the spatial coordinate kth component of the xi related to the ith firefly and d is dimension level. The firefly i movement is attracted to another brighter (appealing) firefly j, which is the relation between the new and old position of firefly i, given by:

9 10 11 12 13

Where, the 2nd term due to attraction and the 3rd term is the randomization with α ∈ [0,∞] being the t randomization parameter, and  i represent a vector of random numbers from a Gaussian distribution/ uniform distribution at time t. Based on (Xin-She Yang, 2008; Yang and He, 2013), the essential steps of the FA can be formulated as the subsequent pseudo code:

xit 1  xit  0e

 rij2

x

t j



 xit   it

(6)

_________________________________________________________________ Firefly Algorithm ____________________________________________________ Begin Determine the objective function f(x), x=(x1......xd )T Generate initial population of fireflies xi (i=1,2,....,n) Determine the light intensity Ii at xi by f(xi) Define the light absorption coefficient γ while (tIi), move firefly i towards j; end if Vary the attractiveness with distance r via exp[-γr] Evaluate new solutions and update light intensity end for j end for i Rank the fireflies and find current global best g. end while Post-process results and visualization. END ___________________________________________________

14

4. Proposed Method

15 16 17 18 19 20 21 22 23 24 25 26 27 28

The proposed system consists mainly of two dependable phases as demonstrated in Figure 1. The first phase includes the image registration process, which required dividing the video under test into frames. Afterward, the binning selection strategy is employed to select the number of frames that will be included in the registration process. To perform the registration, the second phase is involved to determine the optimized values of the scaling factors that used by the velocity field smoothing kernel. The optimization process is achieved via the FA algorithm in the registration framework. To test the proposed system performance, different videos are used and each of them is divided into total 112 number of image frames. To achieve proper distance between the source and the target image, the bin-size must be small. Therefore, the bin-size is set to 14 in order to attain 8 bins, where the total number of 112 frames is to be divided equally, where the total number of bins is (112/14=) 8 that obtained from the target video. Since, the middle frame of the bin in the proposed framework is the seventh frame (of each bin). Thus, the seventh frame is chosen as a fixed image. Subsequently, each bin is allocated as an input and every frame is processed from those bins. The image registration using demon’s algorithm (Khader and Hamza, 2012) is applied using the registration parameters depicted in Table 1. 5

Chakraborty et al.

1 2

Figure 1 Block diagram of the proposed method

3 4 5 6 7 8

The Gaussian low pass filter with integer coefficients is used for velocity field smoothing kernel, thus the total number of iterations used is 200. The integer values of the Gaussian filter guarantee the avoidance of errors that result from floating point operations. The same parameter default values involved for fluid registration in demons [20] is to be used for the proposed system. Thus, the Gaussian filter window size of 60 × 60, with sigma value of 10 pixels for the velocity field smoothing kernel and the value of the Alpha constant value is 2.5 dB, is used.

9 10

Table 1. Registration parameters used in demons registration Total number of iterations

200

11 12 13 14 15 16

Velocity field smoothing kernel

2-D filter

Sigma

Gaussian low pass filter

10 pixels

Alpha (noise) constant

2.5 dB

In the proposed study, the 2-D Gaussian filter for demons registration is employed with optimized velocity field smoothing kernel values using the FA. To optimize the window size (hsize), its values are chosen to vary from 20 to 100, where it is established experimentally that selecting window size less than 20 or greater than 6

1 2 3 4

100 led to erroneous registration. The parameter hsize can be a vector which specifies the number of rows and columns in h. Furthermore, the standard deviation (Sigma) is optimized within the range of 0-20. While, the default value of sigma is 10, where any value more than 20 led to very poor quality for the registered frames. Generally, for both parameters, the distortion increases if the used parameter values are beyond these ranges.

5 6 7 8

Finally, the correlation coefficient obtained by FA is to be used in the demons registration. The highest correlation coefficient is stored with the values of k1, k2 and k3 are used as a fitness function. Once the process is completed, the relationship of the registered image (e') and the original image (e) can be decided using the standard correlation coefficient (Corr) in equation (7), which used to determine the fitness function.

9

  e Corr 

i

ij

- e  f ij - f 

j

(7)

 2  2    eij - e      fij - f     i j  i j 

10 11

Here, f and f' are the transforms of e and e'; respectively. The proposed algorithm can be summed up as follows based Firefly algorithm for optimization: _____________________________________________________________________ The proposed algorithm --------------------------------Begin Objective function Image Registration f(x) x= (k1, k2, k3)T Generate initial population of fireflies xi (i = 1 ,2…..n) Light intensity Icorri at xi is determined by f(xi) (eq. no. 7) Define light absorption coefficient γ while ( t < MaxGeneration ) for i = 1 to n all n fireflies for j = 1 to i all n fireflies if Ireg_corrj > Ireg_corri Move firefly i towards j; end if Attractiveness varies with distance rij via exp [−γ× rij ] Evaluate new solutions and update light intensity end for j end for i Rank the fireflies and find current best. end while Post process on the best so far results and visualization End Image Registration Algorithm: Begin For bin i=1:N, N =total no. of bins For image j=1:M, M = total number of images inside a bin Read the source image; Read the target image; Apply alpha (noise) constant; Apply velocity field smoothing kernel (k1,k2,k3); Begin demons registration; End End for End for End ________________________________________________________________________

12 7

Chakraborty et al.

1 2 3 4 5 6 7 8 9 10 11 12

5. Results Using MATLAB R2012a with 2.20 GHz Intel i3rd generation computer with Microsoft Windows 7 operating system is used to implement the proposed system. The tested videos are obtained from dataset at: (http://www.cipr.rpi.edu/resource/sequences/sequences/sif/ yuv /sif_yuv_tennis.tgz). Meanwhile, the parameters used for FA are: the number of firefly=10, alpha= 0.25; the randomness is 0-1 (highly random), betamn (minimum value of beta) equal 0.20 and gamma (Absorption coefficient) equal 1. In addition, the parameters used for PSO are: inertia = 1.0, correction_factor = 2.0, swarm_size = 10, and the dimension d=3. Furthermore, the used number of iterations in both algorisms is 5, 10, 15, 20, 25, 30, 35, etc. The original image, the source (reference, fixed) image registered, the registered image using the default parameters (velocity smoothing kernels of demons registration) and the registered image using the optimized parameters (velocity smoothing kernels of demons registration) are demonstrated in Figure 2(A), (B), (C) and (D); respectively.

A

B

C

13 14 15 16 17 18 19 20 21 22 23 24 25

26 27 28 29

D

Figure 2 (A) Target image, (B) Source image, (C) Registered image with default parameters, (D) Registered image with optimized parameters (k1, k2 and k3). By visual representation of Figure 2, it is established that involving the FA for optimization has greatly enhanced the registered image quality. The same observed result is proved numerically using two image quality measurement parameters such as structural similarity index (SSIM) and the correlation coefficient. Table 2 illustrated that the correlation has significantly changed using the FA with a value of 0.6108 compared to demons registration with default parameters that provided 0.5591. Similarly, the demons registration based FA for optimization achieved superior structural similarity index (SSIM) value. A SSIM value of 0.4776 is achieved using the FA to attain the optimized parameters, which is greater than the framework without optimization, which realize a value of 0.2736. Table 2. Comparison of optimized and non-optimized framework Comparative parameter

With default parameters

Firefly algorithm (FA)

Correlation

0.5591

0.6108

Structural Similarity index (SSIM)

0.2736

0.4776

Figure 3 illustrated the pair of the moving image and registered image. Figure 3B depicted that the displacement is less significant in the overlapped image of optimized framework than in the default framework (Figure 3A).

8

A

B

1

Figure 3 Overlapped image of moving image with registered image (A) (default parameters, (B) (optimized

2

parameters)

3 4 5 6 7

The Gaussian filter with integer window coefficients is optimized using FA. For the current experimental study, the FA based optimization techniques converged at 25th iteration with population size of 15. The graphical depiction is illustrated in Figure 4 along with the values reported in Table 3. Table 3. Optimization using Firefly algorithm k1 65 72 78

k2 88 82 84

k3 19 20 20

Obtained maximum fitness 0.5914 0.6027 0.6051

20 25 30 35

96 98 98 98

76 98 98 98

20 20 20 20

0.604 0.6108 0.6108 0.6108

Scaling factors

No. of iteration(s) 5 10 15

8 9 10 11 12 13 14 15 16 17 18 19 20 21

No of iteration(s) Figure 4 Optimization of scaling factors using Firefly algorithm Since, the PSO can be considered a good benchmark method. Thus, the results obtained using the FA for optimizing demons registration is compared to using the PSO algorithm. The comparison is conducted using the same Gaussian filter. Thus, the obtained FA results reported in Table 3 are compared to those attained using the PSO that reported in Table 4.

9

Chakraborty et al.

1

No. of iteration(s)

k1

k2

k3

Obtained maximum fitness

5

93

98

20

0.6099

10 15 20

98 98 93

84 98 98

20 20 20

0.6082 0.6108 0.6098

25 30 35

95 98 98

81 98 98

20 20 20

0.6081 0.6108 0.6108

The results obtained in Table 3 and Table 4 are represented in Figure 5 to compare both results, which supported the proposed claim.

Fitness

2 3 4

Table 4. Optimization using Particle Swarm Optimization

5 6

No of iteration(s)

7

Figure 5 Comparative study of optimized fitness values using PSO and FA

8 9 10 11

Figure 5 clarified that the PSO is converged after 30th iteration with the same population size of 15, claiming that FA is superior technique for demons registration optimization compared to the PSO. Table 5 and Figure 6 demonstrated the time comparison of using the FA and the PSO based optimization framework for registration.

12

Table 5. Time complexity of FA versus the Particle Swarm Optimization No. of Iteration(s) 5 10 15 20 25 30 35

Firefly Algorithm (sec) 2141.91 3748.07 12374.13 15914.34 19820.22 23545.22 39581.19

10

Particle Swarm Optimization(sec) 6871.9 12599.6 18567 32517.4 43158.7 57891 75671.6

Time (sec)

1 2

No of iteration(s)

3

Figure 6 Time complexity comparison of Firefly Algorithm and Particle Swarm Optimization

4 5 6 7 8 9 10 11 12 13

Visual representation of the graph (Figure 6) and comparative study from Table 5 proved that the involving FA is faster than that of the PSO. Consequently, from the above results it is established that the proposed registration approach enhanced the image quality more than the classical registration. In addition, the comparative study of the FA versus PSO proved that FA based optimization framework is more stable and produced superior results than the PSO based optimization framework. Besides, the FA based optimization framework is faster than PSO based optimization framework. Therefore, the overall conclusion proved that the FA based optimization for demons registration is more stable as well as faster than PSO based optimization framework.

14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29

The current work proposed a method for registering video sequences based on demons registration based FA for optimization. The suggested framework employed the Gaussian filter values for velocity smoothing kernel scaling factors (k1, k2, k3 ) to be optimized. The FA used the correlation coefficient as the fitness function, where the highest fitness value was stored in order to use the corresponding scaling factor values for post-processing. The performance of the proposed FA based framework was compared to the classic demons algorithm with default parameters without optimisation, which established the superiority of the proposed system. Moreover, comparing the proposed FA based optimization framework to the PSO based optimization framework demonstrated that the fitness value, scaling factors were higher in the FA than PSO. Additionally, the FA achieved faster convergence with 25 iterations than the PSO, which converged at 30 iterations. Subsequently, demons registration based FA is a faster framework than using the PSO.

30

References

31 32 33 34 35 36

[1] Bulsara, V., Bothra, S., Sharma, P. and Rao, K.M.M. (2011) Low Cost Medical Image Processing System for Rural/Semi Urban Healthcare. IEEE Conference Publications on Recent Advances in Intelligent Computational Systems (RAICS), 724 – 728. [2] Dey, N., Roy, A. B. and Das, A. (2012a) Optical Cup To Disc Ratio Measurement In Glaucoma Diagnosis Using Harris Corner. Third International Conference Computing Communication and Technologies 2012(ICCCNT12), 2012, Coimbatore, 1-5.

6. Conclusion

Acknowledgement This work was co-funded by European Union through European Regional Development Funds Structural Operational Program “Increasing of Economic Competitiveness” Priority axis 2, operation 2.1.2. Contract Number 621/2014.

11

Chakraborty et al.

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55

[3] Dey, N., Biswas, S., Roy, A. B., Das, A. and Chaudhuri, S. S. (2012b) Analysis of Photoplethysmographic Signals Modified by Reversible Watermarking Technique using Prediction-Error in Wireless Telecardiology. International Conference of Intelligent Infrastructure , 47th Annual National Convention of CSI –Kol, 01 - 02 December 2012, 1-5. [4] Sharma, T. K. and Pant, M. (2013) Enhancing different phases of artificial bee colony for continuous global optimisation problems. Int. J. of Advanced Intelligence Paradigms 5(1/2),103 – 122. [5] Roy, P., Chakraborty, S., Dey, N., Dey, G., Ray, R. and Dutta, S. (2014) Adaptive Thresholding: A comparative study. International Conference on Control, Instrumentation, Communication and Computational Technologies-2014, 10-11 July 2014, 1182-1186. [6] Jakšić, B., Gara, B., Petrovic, M., Spalevic, P. and Lazic, L. (2015) Analysis of the Impact of Front and Back light on Image Compression with SPIHT Method during Realization of the Chroma Key Effect in Virtual TV Studio. Acta Polytechnica Hungarica 12(2). [7] Mathew, S. P., Balas, V. E., Zachariah, K. P. and Samuel, P. (2014) A Content-based Image Retrieval System Based on Polar Raster Edge Sampling Signature. Acta Polytechnica Hungarica 11(3). [8] Pal, G., Acharjee, S., Rudrapaul, D., Ashour, A. S. and Dey, N. (2015) Video Segmentation using Minimum Ratio Similarity Measurement. International Journal of Image Mining 1(1), 87-94. [9] Nandi, D., Ashour, A. S., Samanta, S., Chakraborty, S., Salem, M. and Dey, N. (2015) Principal Component Analysis in Medical Image Processing. A Study, International Journal of Image Mining 1(1), 1-17. [10] Gospodinova, E., Dey, N., Domuschiev, I., Ashour, A. S., Sifaki-Pistolla, D. and Gospodinova, M. (2015) Analysis of Heart Rate Variability by Applying Nonlinear Methods with Different Approaches for Graphical Representation of Results. International Journal of Advanced Computer Science and Applications 6(8), 38-44. [11] Chowdhury, S., Chakraborty, S., Karaa, W., Ray, R. and Dey, N. (2014) Effect of Demons Registration on Biomedical Content Watermarking. International Conference on Control, Instrumentation, Communication and Computational Technologies 2014, 509-514. [12] Araki, T., Ikeda, N., Dey, N., Chakraborty, S., Saba, L., Kumar, D., Godia, E., Jiang, X., Gupta, A., Radeva, P., Laird, J., Nicolaides, A. and Suri, J. (2015) A comparative approach of four different image registration techniques for quantitative assessment of coronary artery calcium lesions using intravascular ultrasound. Computer Methods and Programs in Biomedicine 118(2),158-172. [13] Crum, W.R., Hartkens, T. and Hill, D.L.G. (2004) Non-rigid Image Registration: Theory and Practice. The British Journal of Radiology 77, 140–253. [14] Thirion, J.P. (1998) Image Matching as a Diffusion Process: An Analogy with Maxwell’s Demons. Medical Image Analysis 2, 243–260. [15] Brown, L. (1992) A Survey of Image Registration Technique. ACM Comput. Surv 24, 325-376. [16] Rouet, J. M., Jacq, J. J. and Roux, C. (2000) Genetic algorithms for a robust 3-D MR-CT registration. IEEE Trans. Inform. Techno. Biomed 4, 126–136. [17] Wachowiak, M. P., Smolíková, R., Zheng, Y., Zurada, J. M. and Elmaghraby, A. S. (2004) An Approach to Multimodal Biomedical Image Registration Utilizing Particle Swarm Optimization. IEEE Transactions on Evolutionary Computation 8, 289-302. [18] Klein, S., Staring, M. and Pluim, J.P.W. (2007) Evaluation of Optimization Methods for Nonrigid Medical Image Registration Using Mutual Information and B-Splines. IEEE Transactions on Image Processing, 16(12), 2879-2891. [19] Chen, Y.-W., Lin, C.-L. and Mimori, A. (2008) Multimodal Medical Image Registration Using Particle Swarm Optimization. Proc. of ISDA2008, 3, 127-131. [20] Mohamed, W. and Hamza, A. (2010) Medical image registration using stochastic optimization. Optics and Lasers in Engineering 48(12), 1213–1223. [21] Meskine, F., Mezouar, M. C. E. and Taleb, N. (2010) A Rigid Image Registration Based on the Nonsubsampled Contourlet Transform and Genetic Algorithms. Sensors 10(9), 8553–8571. [22] Zheng, L. and Tong, R. (2011) Image Registration Algorithm Using an Improved PSO Algorithm. ICCIC 2011, Part IV, CCIS, 234, 198–203. [23] Singhai, R. and Singhai, J. (2012) Registration of Satellite Imagery using Genetic Algorithm. Proc of the World Congress on Engineering, WCE 2012, London, UK, Vol. 2, pp. 201-206. [24] Bejinariu, S.-I., Rotaru, F., Nita, C., Luca, R. and Costin, H. (2013) Parallel image registration using bioinspired computing. in E-Health and Bioengineering Conference (EHB), 2013 , pp.1-4. 12

1 2 3 4 5 6 7 8 9 10 11 12 13 14

[25] Hwuang, E., Danish, S., Rusu, M., Sparks, R., Toth, R. and Madabhushi, A. (2013) Anisotropic Smoothing Regularization (AnSR) in Thirion’s Demons Registration Evaluates Brain MRI Tissue Changes Post-Laser Ablation. 35th Annual International Conference of the IEEE EMBS Osaka, Japan, 3 - 7 July, pp. 4006-4009, 2013 [26] Caspi, Y., Simakov, D. and Irani, M. (2006) Feature-based sequence-to-sequence matching. International Journal of Computer Vision 68, 53–64. [27] Scott, A. M., Khare, R., Bascom, R. and Higgins, W. E. (2013) Interactive CT-Video Registration for the Continuous Guidance of Bronchoscopy. IEEE Transactions on Medical Imaging 32(8), 1376-1396. [28] Yang, X. S. (2008) “Nature-Inspired Metaheuristic Algorithms”, Luniver Press, UK. [29] Yang, X. S. and He, X. (2013) Firefly Algorithm: Recent Advances and Applications. Int. J. Swarm Intelligence 1(1), 36-50. [30] Khader, M. and Hamza, A.B. (2012) An information-theoretic method for multimodality medical image registration. Expert Systems with Applications 39(5),5548–5556.

15 16 17

13