You are on page 1of 6

FSBee: An ABC-based Filter Sequencing for Image Denoising

Gerard Wilfred Jusayan drareg_jusayan@yahoo.com John Paul T. Yusiong jpyusiong@gmail.com

Division of Natural Sciences and Mathematics University of the Philippines Visayas Tacloban College Tacloban City Abstract: Image denoising has long been a field of interest of many researches due to its importance in the field of image processing. Among the numerous approaches in solving this problem, the most common is the use of smoothing filters. The roster of available filters that can be used is very broad, each of them distinct and may be effective against certain types of noise only. This limitation can be surpassed by designing an approach that can restore an image regardless of the type of noise it has, and this can be realized by collectively using various filters and apply them in a sequence. The generation of this sequence must be governed by an optimization algorithm to ensure that the filter sequence would yield an optimal output, and Artificial Bee Colony has been employed in this study to perform this task, taking PSNR as its fitness function. Experiments performed in this research proved that ABC is an effective optimization tool, and produced outputs whose quality could compete to a recently published research using the same approach. Keywords: Image Denoising Problem, Artificial Bee Colony, Filter Sequence, PSNR

1 Introduction
Images have long been part of the lives of the people. Its functions range from leisure, such as capturing pictures to cherish important or significant events, to practicality, in the case of surveillance imaging for keeping the security of an establishment. These are only some of the functions from the vast applications that images play in our daily lives, almost becoming a commodity that people would find hard to live without. But images, specifically digital images, dont always come unchallenged. Noise is the most common perpetrator that corrupts digital image quality. The various types of noises acquired by images are caused by various factors. The presence of the variety of noise also allowed a diverse collection of algorithms that mostly serve as solutions to certain types of noise, only few among these algorithms can be called versatile. The existence of a study which makes use of filters, one of the most conventional image denoising methods, combined with Genetic Algorithm (GA) optimization in generating a sequence of filters to be applied to the noisy image [2], encouraged the researcher to propose FSBee, an Artificial Bee Colony-based Filter Sequencing for Image Denoising. Aside from GA, several swarm-intelligence optimization algorithms have been successful in this task. These are Ant Colony Optimization (ACO) [5] and Particle Swarm Optimization (PSO) [4], although both of these optimization algorithms were employed on a different approach. The rest of the paper can be followed as such: Section 2 will define and describe the Image Denoising Problem. Section 3 will highlight the optimization algorithm used in this study, ABC. The proposed approach using ABC and filter sequence would be discussed in Section 4. Section 5 and 6 will cover results and discussions and conclusion respectively.

2 The Image Denoising Problem


Although it is impossible to restore a noisy image to its original quality, the goal of image denoising is to produce a denoised image whose quality is as close to the original. Formally, given a noisy version of an image, , the objective of restoration is to obtain an estimate, , of the original image, u. The denoised estimate is wanted to be as close as possible to original. Additionally, the more one knows about the noise, the closer the denoised estimate, , will be to u [1].

(a)

(b)

(c)

Figure 1: (a) Sample Image (b) Noisy Image (c) Restored Image Image denoising is an important pre-processing step for image analysis, so the challenge that lies in devising an efficient and reliable approach to denoising images regardless the type of noise and retaining the images distinctive features is a considerable one for a better and more accurate analysis of a given digital image.

3 Artificial Bee Colony Optimization Algorithm


Artificial Bee Colony (ABC) is a relatively new swarm-based metaheuristic algorithm compared to others which have been circulating the scholastic community since the early nineties. It was introduced in 2005 by Dervis Karaboga in his research entitled An idea based on honey bee swarm for numerical optimization. ABC emulates the bee colonys food exploration and exploitation activities in finding its solutions. It was originally proposed to answer the need for optimizing numerical problems [3], and now it has been employed for a wide variety of problems ranging from theoretical mathematics to real world applications. There are three important roles in the implementation of ABC, namely employed bees, scout bees and onlooker bees. Employed bees are those who have found rich food sources. Employed bees represent feasible solutions [3]. Onlooker bees are unemployed bees that stay in the hive and wait for the employed bees. They follow employed bees that have successfully found good solutions, and search for neighbor solutions of the employed bee that they followed. In case the neighbor solution that the onlooker has found is more promising, the onlooker bee takes the place of the employed bee that it followed [3]. Scout bees are bees that search for a neighbor solution to a solution that cannot be improved for a certain number of cycles [3].

The ABC process is composed of these major phases that keep on repeating until the termination criteria have been met: 1. Initialization Assign random solutions to employed bees 2. Employed Bees Phase All employed bees will try to improve their current solutions by performing a neighborhood search 3. Onlooker Bees Phase All onlooker bees will try to find better solutions than the employed bee that they followed by performing a neighborhood search 4. Scout Bees Phase Look for solutions that hasnt been improved for a certain amount of cycles and perform neighborhood search in this stale solution 5. Memorize Best Solution Keep track of the bee with the best solution

Figure 2: ABC Flowchart

4 FSBee: Solving the Image Denoising Problem using ABC


Each bee is a sequence of filters which will be represented as an array with a fixed size equal to the number of filters allowed to be used. The values of the array range from 0 to 7, and may be repeated in the sequence. Respective values represent a certain filter that would be added to the sequence, enumerated as follows: 0 no filter, 1 Mean Filter, 2 Median Filter, 3 Mode Filter, 4 Circular Filter, 5 Pyramid Filter, 6 Cone Filter and 7 Gaussian Filter. Standard employed bees population will be equal to the half of the onlooker bees population. There will only be one scout bee. It is ABCs goal to find the best solution for our problem in the form of the fittest bee. And for us to be able to calculate the fitness of each bee, we will refer to Equation (1). fiti= PSNRi (1) With our fitness function, Equation (1), the higher the fitness value for each bee, the better solution they are. Peak Signal to Noise Ratio (PSNR), a standard performance index for image quality, is the best metric for our bees fitness. PSNR is described in the next page:
3

PSNR= 1010

255 2 , ,
2

(2)

where x is the original noise-free image, y being the restored image, M and N the sizes of x and 0iMand 0jN. The quality of the restored image is directly proportional to the PSNR value.

Figure 3: Overall FSBee program flow ABC will repeatedly search for the best filter sequence until we have met the termination criterion which in this case is the maximum number of cycles.

5 Experiments and Results


In order to prove ABCs effectiveness in solving the image denoising problem, experiments were setup to accommodate all combination of the noise to be added to Circuit. Experiments were also conducted on several standard test images that have also been used in [2]. The effects of varying the standard ABC parameters were also observed through other experiment setups. To ensure statistical adequacy in the results, the experiments to analyze the effects of varying the ABC parameters were prepared for 30 runs. The standard ABC parameters are defined in Table 1 below: Table 1: Standard ABC Parameter Values EB OB Scout Size Parameter 10 20 1 Value Table 2: Different noise types applied to Circuit Setup C1 C2 C3 C4 C5 C6 C7 C8 A B C D Table 2 shows the different experiment setups used to test the effectiveness of ABC. Setup with shaded cells on A means that Gaussian noise with = 25 will be added to the image, B for Poisson noise with = 25, C for Speckle noise with = 25 and D for Salt and Pepper noise with probability = 5%.
4

Search Limit 5

Max Cycle 50

Kernel Size 33

C9

C10 C11 C12 C13 C14 C15

30 28 PSNR Value 26 24 22 20 18 16 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 Run

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11 C12 C13 C14 C15

Figure 4: Results on different experiment setups of Circuit Figure 4 graphs the results of performing 30 runs of optimization to the different test cases that we have defined in Table 2. The quality of the resulting image can be observed as inversely proportional to the levels of noise it has to reduce.

(a)

(b)

(c)

Figure 5: (a) Circuit (272280) (b) Noisy Image (Salt and Pepper 5%) (c) Restored Image Figure 5.c shows the resulting image when restoration is performed on Circuit, one of the test images used in [2], on standard ABC parameters. Figure 6 below graphs the result that we have acquired with the same experiment that we have conducted in Figure 5. It is evident that ABC was able to find the global best solution in relatively early cycles.

Figure 6: PSNR Values acquired through standard ABC parameters on Circuit


5

Figure 7: Result when OB size varies

Figure 8: Result when EB size varies

On the 30 experimental runs, the average fitness value acquired is higher when the bee sizes are high, as shown in Figure 7 and Figure 8. Higher solution values were acquired when the Employed Bee as well as Onlooker Bee population was high since having more bees allows a greater chance for ABC to find a mutate or neighbor of our fittest bee which can yield better fitness.

6 Conclusions
This paper has proven that ABC is an effective optimization algorithm for coming up with a filter sequence that can restore a noisy image, as experimental results can attest. It could also be observed that increasing the bee population helps in acquiring a higher average fitness on all the experimental setups.

REFERENCES
[1] [2] Ghazel, M.: Adaptive Fractal and Wavelet Image Denoising (Unpublished) (2004) Gupta, S., Kumar, R., Panda, S-K.: A Genetic Algorithm Based Sequential Hybrid Filter for Image Smoothing. International Journal of Signal and Image Processing vol. 1 no 4 (2010) 242 - 248 Karaboga, D.: An idea based on honey bee swarm for numerical optimization. Technical Report- TR06, Erciyes University, Engineering Faculty, Computer Engineering Department (2005) Roomi, S-M-M., Muthu Karuppi, P.L., Rajesh, P. and Guru Revathi, B.: A Particle Swarm Optimization Based Edge Preserving Impulse Noise Filter. Journal of Computer Science vol.6 no 9 (2010) 1014 1020 Tian, J., Yu, W. and Ma, L.: AntShrink: Ant colony optimization for image shrinkage. In Proceedings of Pattern Recognition Letters vol. 31 (2010) 1751-1758

[3]

[4]

[5]

You might also like