OPTIMIZING SVM FOR IMAGE RANKING USING ENHANCED ABC

Transcription

1 OPTIMIZING SVM FOR IMAGE RANKING USING ENHANCED ABC NIDHI GONDALIA 1, NIRALI MANKAD 2 1 PG student M.E. Cop., Noble Group of Institutions, nidi.gondalia28@gail.co 2 C.S.E. DEPT., Noble Group of Institutions, nirali.ankad@gail.co Abstract It is saying that iage is worthwhile 100 words. It is better way to explain anything using iage. Here in this paper I a proposing an enhanced swar based optiization algorith to optiize iage ranking procedure. I a using Support Vector Machine for iage ra nking procedure. Swar based technique works on intelligent group work of eber of particular swar. There are any swar based techniques are available now a days like ACO, PSO, ABC etc. ABC is working on intelligent behavior of honey bee swar. To reove soe of liitation of ABC algorith here I hybridize ABC with Genetic algorith. SVM is good classifier and by optiization process of weight vector we can get better perforance of it. In this paper, we provide a thorough and extensive overview of ost research work focusing on the application of ABC, with the expectation that it would serve as a reference aterial to both old and new, incoing researchers to the field, to support their understanding of current trends and assist their future research prospects and directions. Also new proposed architecture of Enhanced ABC algorith, coparison between results of ABC and EABC for iage ranking is also given here. Keywords : Artificial Bee Colony Algorith, Iage Ranking, SVM, Swar based techniques I. INTRODUCTION Optiization is the process to get optiu output with iniu or current input [1]. Iage ranking is the process of classifying different iages based on their content. In real word any ties we search iage by words it s little bit cubersoe to find any iage by word on internet because we can t explain any iage by phrase of few words. To get exact result we can search iage by its content and it leads us to iage based retrieval. For iage based retrieval we need to rank iage based on its content. In recent are any iage ranking algorith are available like SVM, ANN, Naive Bayes Classifier, k-nearest Neighbor, Decision Trees etc. Fro all this algorith SVM has high generalization capacity so it can correctly classify iages with sall set of training exaples. Support Vector Machine is proposed by vapnic in SVM works based on statistical learning theory. Statistical learning theory is not only a tool for the theoretical analysis but also a tool for creating practical algoriths for pattern recognition.[7]. Many optiization techniques are available like Evolutionary algoriths, swar based algoriths. I concentrate on swar based techniques to optiize SVM. All swar based techniques works based on ain four principle which are: I) Positive feedback, II) negative feedback, III) fluctuations and IV) ultiple interactions. Here I a working on artificial bee colony algorith which works based on collective and clever behavior of honey bee swar. The reaining parts of the paper are organized as follows: Coplete ABC algorith is briefly outlined in Section II, iage ranking algorith is discussed in Section III. Section IV briefly explains proposed architecture of our EABC algorith and finally in section V we will copare result of both ABC and EABC to optiize SVM.

2 II. THE ABC ALGORITHM a) Forging behavior of bees in nature In nature honey bees works on rule divide and concur. All the work of nest is divided between group of bees. All bees are divided in ain three groups scout bee, onlooker bee and honey bee. Workings of all three groups are given below: Figure 1: Intelligent behaviour of honey bee foraging 1) Scout bee: they start searching of food sources randoly without any inforation about the food source. 2) Eploy bee: when scout bee finds any food source then it reebers inforation about that source and collects nectar fro that source and also shares inforation about that source to other bees by perforing waggle dance. 3) Onlooker bee: those bees watch the waggle dance perfored by onlooker bee and then goes to the food source found by eploy bee to collect the nectar. 4) Waggle Dance: this is the special type of oveent perfored by the eploy bee to share the inforation about the food source with onlooker bees. b) Matheatical odel of ABC The ABC consists of four ain phases: 1. Initialization Phase: The food sources, whose population size is SN, are randoly generated by scout bees. Each food source, represented by is an input vector to the optiization proble, has D variables and D is the diension of searching space of the objective function to be optiized. The initial food sources are randoly produced via the expression (1)

3 x l rand 0,1 * u l. i i i...(1) Where and are the upper and lower bound of the solution space of objective function, rand (0, 1) is a rando nuber within the range [0, 1]. 2. Eployed Bee Phase: Eployed bee flies to a food source and finds a new food source within the neighborhood of the food source. The higher quantity food source is eorized by the eployed bees. The food source inforation stored by eployed bee will be shared with onlooker bees. A neighbor food source is deterined and calculated by the following equation (2) i i i i ki v x x x.....(2) Where i is a randoly selected paraeter index, is a randoly selected food source, is a rando nuber within the range [-1, 1]. The range of this paraeter can ake an appropriate adjustent on specific issues. The fitness of food sources is essential in order to find the global optial. The fitness is calculated by the following forula (3), after that a greedy selection is applied between x and f. 1,fx 0 1 fx fit x 1 f x,f x 0....(3) Where is the objective function value of. 3. Onlooker Bee Phase: Onlooker bees calculates the profitability of food sources by observing the waggle dance in the dance area and then select a higher food source randoly. After that onlooker bees carry out randoly search in the neighborhood of food source. The quantity of a food source is evaluated by its profitability and the profitability of all food sources. P is deterined by the forula p fit 1SN x fit x....(4) Where fit (x )is the fitness of x. Onlooker bees search the neighborhood of food source according to the expression (5) v x x x.. i i i i ki...(5) 4. Scout Phase: If the profitability of food source cannot be iproved and the ties of unchanged greater than the predeterined nuber of trials, which called "liit", the solutions will be abandoned by scout bees. Then, the new solutions are randoly searched by the scout bees. The new solution will be discovered by the scout by using the expression (6)

4 x l rand 0,1 * u l i i i (6) rand 0,1 is a rando nuber within the range [0,1], and are the upper and lower bound of the solution space of objective function.[5] c) Merits & Deerits of ABC Merits 1. ABC is flexible so can easily odified and hybridize with other algorith.[3] 2. ABC is quite robust and not need to give any starting point.[3] 3. The ABC algorith eploys distributed coputation, which avoids preature convergence.[4] Deerits 1. There is no centralized processor to guide the bees syste towards good solutions[5] 2. This algorith has slower convergence than other heuristic-based algoriths.[6] 3. There does not exist any standard software for the ABC algorith as well.[1] III. PROPOSED ARCHITECTURE OF OUR EABC ALGORITHM By analysis of different optiization technique it is clear that ABC overcoes any liitations of current optiization techniques. And fro one of the advantage of ABC is that we can also reove liitation of ABC by odifying it. One of the drawback of ABC is that it stuck at local optia and we can t get global optial solution. To overcoe this drawback we hybridize ABC with genetic algorith. Here architecture of proposed syste is given below: Figure 2 : Architecture of Proposed Miner

5 International Journal of Advance Engineering and Research Developent (IJAERD) Volue 1,Issue 5,May 2014, e-issn: , print-issn: Stage 1: Input Raw Iages This is the very first stage where the input is fed to the syste. For ining proble we have considered cobination of iages fro real world. Main purpose of our proposed syste is to rank all iages that fro which class particular iage belongs. Stage 2: Foratted Vectors At this basic step histogra of iage is build fro iage using key features and diensions of iage. For this process first we use SIFT features of iage. Then this vector quantized into visual words. The frequency of each visual words is then recorded in a histogra. And final vector for iage is a cobination of these histogras. Stage-3 Training iages Our EABC iner is supervised iner so we have to give soe training iages to train our iner. For training process we provide set of iages fro any particular class and also we can separate particular iages in raw iages by.at file. Stage 4- Enhanced ABC Algorith to optiize SVM classifier This is the heart of our syste. Entire process of ranking starting fro training to validation is done at this stage. Main work of our EABC optiization algorith is to find out optial feature vector. Stage-5 Training iages A validation set is required to validate our ranking results. This iage set is directly given fro raw data which contains cobinations of all class iages. It is very uch alike the training set but ain difference is that it contains all class iages and our iner provides resulted set of iages which is fro sae class. Stage 6- Output Ranked Iage set & Accuracy check This is the output of our iner. Here we display possible 64 iages fro our raw data which belongs to sae class. Here we also calculate that how any iages are correctly retrieved and based on that calculation accuracy is calculated of our syste. IV. COMPARISON OF RESULT BETWEEN ABC & EABC To verify the effectiveness of the ABC algorith, we conducted experients on different types of class fro various real doains. The ai behind using different class iages for experientation is to prove the consistency of the SVM-EABC classifier in different doains. Here I have given description of two datasets which I a using in this All rights Reserved 5

6 International Journal of Advance Engineering and Research Developent (IJAERD) Volue 1,Issue 5,May 2014, e-issn: , print-issn: Table 1 : PASCAL Dataset description Table 2 : Caltech-101 Dataset Description PASCAL VOC 2007 benchar k Caltech-101 benchark dataset Class Nuber of Iages airplane 238 Class Nuber of Iages Person 2008 Motorbike 245 Car 713 Horse 287 Background 1455 anial 356 Person 1692 Bird 856 Car 954 Background 1576 V. RESULT ANALYSIS The paraeters of the ABC algorith were set to the sae value for all the dataset probles: Colony size (NP = 20), nuber of food sources NP/2, liit = 100 and the axiu nuber of cycles was MCN = Here result of our SVM-ABC classifier is copared with pure SVM classifier for different class iages and for two benchark iage datasets. Coparison between pure SVM and SVM-ABC is given using chart representation for different class iages. 70% 60% 50% 40% 30% 20% SVM-ABC SVM-EABC 10% 0% Aeroplane Horse Person Motorbike All rights Reserved 6

7 International Journal of Advance Engineering and Research Developent (IJAERD) Volue 1,Issue 5,May 2014, e-issn: , print-issn: Figure 3 : coparison of perforance of SVM-ABC and SVM-EABC on PASCAL dataset Figure 4 : coparison of perforance of SVM-ABC and SVM-EABC on Caltech-101 dataset Fro result graph it is clear that SVM-EABC gives better perforance than SVM-ABC. VI. CONCLUSION In this paper, variety of research article in the doain of ABC has studied. Fro the in depth literature survey, it is observed that a large part of research was concentrated towards odifying the ABC algorith to solve a variety of probles, including engineering design probles, scheduling probles, data iing probles etc. Although ABC has great potential, it was clear to the scientific counity that soe odifications to the original structure are still necessary in order to significantly iprove its perforance. Further, in this paper I have given a swar based Enhanced ABC approach for optiizing SVM iage ranking. By analysis of result of experients it is clear that EABC is ore proising than ABC. As a future work we can also ipleent EABC for reote sensing and different type of iages. VII. ACKNOWLEDGMENT The satisfaction and euphoria that accopany the successful copletion of any task would be incoplete without the ention of the people who ade it possible, whose consistent guidance and encourageent crowned y efforts with success. Nidhi Gondaliya highly thankful to y guide Nirali Mankad without whose guidance the work would not have been aterialized and he helps e to solve y difficulties arising during this report preparation. Also N.B. Gondaliya would like to acknowledge all other faculty eber of Coputer Departent and y classates who have coe forward to help e in every way at any tie. With all this assistance little for reains for which I can take credit. REFERENCES [1]. Karaboga D., Basturk B., A powerful and efficient algorith for nuerical function optiization: Artificial bee colony (ABC) algorith, Journal of Global Optiization 39, , All rights Reserved 7

8 International Journal of Advance Engineering and Research Developent (IJAERD) Volue 1,Issue 5,May 2014, e-issn: , print-issn: [2]. Fahad S. Abu-Mouti, Mohaed E. El-Hawary Overview of Artificial Bee Colony (ABC) Algorith and Its applications, IEEE 2012 [3]. R.S. Rao, SVL Narasiha, and M. Raalingaraju. Optiization of distribution network configuration for loss reduction using artificial bee colony algorith. International Journal of Electrical Power and Energy Systes Engineering, 1(2): , [4]. Beatriz A. Garro, Huberto Sossa, Roberto A. V azquez, Artificial Neural Network Synthesis by eans of Artificial Bee Colony (ABC) Algorith, /11/$ IEEE [5]. Karaboga, D. and Basturk, B., On the perforance of artificial bee colony (ABC) algorith, App lied Soft Coputing, 8, , [6]. Basturk, B. and Karaboga, D., An artificial bee colony (ABC) algorith for nueric function optiization, Proceedings of the IEEE Swar Intelligence Syposiu, Indianapolis, USA, [7]. Zhi-Gang Fan, Kai-An Wang, Feature Selection for Fast Iage Classification with Support Vector Machines, Springer-Verlag Berlin Heidelberg 2004 [8]. Baykasoglu, L. Ozbakir, and P. Tapkan. Artificial bee colony algorith and its application to generalized assignent proble. Swar Intelligence: Focus on Ant and Particle Swar Optiization, 2007, [9]. Haiyan Quan and Xinling Shi. On the analysis of perforance of the iproved artificial-bee-colony algorith, in natural coputation, 7, 2008, [10]. H. Narasihan. Parallel artificial bee colony (pabc) algorith. In Nature & Biologically Inspired Coputing, World Congress on, 2009, [11]. P.W. Tsai, J.S. Pan, B.Y. Liao, and S.C. Chu. Enhanced artificial bee colony optiization. International Journal of Innovative Coputing, Inforation and Control, 2009, 5(12). [12]. B. Akay and D. Karaboga. A odified artificial bee colony algorith for real paraeter optiization. Inforation Sciences, [13]. M. Sonez. Artificial bee colony algorith for optiization of truss structures. Applied Soft Co puting, 11, 2010, [14]. R. Luo, T.S. Pan, P.W. Tsai, and J.S. Pan. Parallelized artificial bee colony with ripple-counication strategy. In Genetic and Evolutionary Coputing (ICGEC), 2010 Fourth International Conference on, IEEE, 2010, [15]. Wenping Zou, Young Zhu, Hanning Chen, Xin Sui, A clustering approach using cooperative artificial bee colony algorith, Hindawi publishing corporation, 21(459796), 2010, [16]. D. Karaboga and B. Akay. A odified artificial bee colony (ABC) algorith for constrained optiization probles. Applied Soft Coputing, 11, 2011, [17]. Tarun Kuar, Millie Pant, V.P. Singh, Artificial bee colony algorith with self-adaptive colony size, proceeding of the 2nd international conference on swar, evolutionary and etic coputing, 1, 2008, [18]. Bin Wu and Chu hua Qian. Differential artificial bee colony algorith for global nuerical optiization, in journal of coputers, 6(5), 2011, [19]. M. El-Abd. Opposition-based artificial bee colony algorith. In Proceedings of the 13th annual conference on Genetic and evolutionary coputation, ACM, All rights Reserved 8