Feature coding for image classification based on saliency detection and fuzzy reasoning and its application in elevator videos

Transcription

1 Feature codng for mage classfcaton based on salency detecton and fuzzy reasonng and ts applcaton n elevator vdeos Xao Lv *, Dngdong Zou, Le Zhang and Shangyuan Ja Chongqng specal equpment nspecton and research nsttute No.5 Furongyuan Road Northern New Chongqng PEOPLE S REPUBLIC OF CHINA lvxao87@6.com, cq_zdd@6.com, zl_86@63.com, @qq.com Abstract: - Feature codng s an fundamental step n bag-of-words based model for mage classfcaton and have drawn ncreasng attenton n recent works. However, there stll exts ambguty problem, and t s also senstveness to unusual features. To mprove the stablty and robustness, we ntroduce salency detecton and fuzzy reasonng rules to propose an novel codng scheme. In detal, salency maps generated by salency detecton are frst used to dvde each mage nto salent and non-salent regon, then a structured dctonary s obtaned by combng two separated codebooks n them. Secondly, fuzzy reasonng rules are ntroduced to choose the most salent and stable codewords to encode. Fnally, salency maps are ncorporated nto poolng operaton named salency based spatal poolng to ntroduce spatal nformaton. Experments on several datasets demonstrate our approach outperforms all other codng methods n mage classfcaton. Furthermore, we also apply t nto elevator vdeo event classfcaton, whch shows the potental applcaton n ntellgent elevator vdeo survellance, such as overload detecton, volence detecton, vdeo summarzaton. Key-Words: - Image classfcaton, feature codng, salency detecton, fuzzy reasonng, elevator vdeo event Introducton Automatc mage classfcaton s one of the most fundamental problems n computer vson and pattern recognton, whose am s to assgn one or more category labels to an mage. It has drawn ncreasng attenton from the researchers around the world due to ts wdespread prospects n a wde range of applcatons, e.g., mage retreval [, 36], vdeo retreval [], vdeo survellance [3], humancomputer nteracton [4], web content analyss [5], and bomedcal [6, 37]. There are many approaches proposed for mage classfcaton n the lteratures. Among them, the bag-of-words (BOW) model [7] and ts extensons [8] acheve the state-of-the-art performance n several famous databases, such as Caltech 0 [9], Scenes 5 [0], Caltech 56 [], and PASCAL VOC []. The BOW quantzes local descrptors nto dscrete vsual codewords and counts ther occurrence frequences n the entre mage. Then the resultng hstogram s used as the mage representaton. Fg. shows the general framework of the BOW model. It usually comprses of the followng common steps: () Feature extracton. It extracts mages local features by detectors or dense samplng and then calculates ther descrptors, such as Harrs detector [3], affne nvarant salent regon detector [4], SIFT (Scale-Invarant Feature Transform) [5] descrptor, HOG (Hstogram of Orented Gradent) [6] descrptor. () Codebook generaton. After obtaned local descrptors, a codebook s usually needed to represent them. It s typcally generated by clusterng (e.g., K-means [7]) over a subset of descrptors, whch s randomly sampled from all descrptors n database n real applcaton for computatonal effcency. (3) Feature codng and poolng. In ths step, each local descrptor frst actvates a number of codewords, and generate a codng vector. Then, all responses on each codeword are ntegrated nto one value by feature poolng. Varous codng and poolng strateges wll be descrbed n detal n Secton. The output of ths step s a vector whose length s equal to the sze of the codebook, namely the fnal mage representaton. (4) Classfcaton. Fnally, the mage representaton vectors are sent to a classfer, such as SVM (Support Vector Machne) [8-9] for classfcaton. Among these steps, feature codng and poolng s the fundamental component, whch wll greatly nfluences mage classfcaton n terms of both * Correspondng author. E-ISSN: Volume 3, 04

2 Feature Extracton Codebook Generaton Feature Codng and Poolng Classfcaton Butterfly Fg. The framework of the Bag-of-Words based model. accuracy and computaton cost [0]. Thus, t has drawn ncreasng attenton n recent works, and varous good strateges have been proposed n the lteratures. Hard assgnment (HA) [7-8] s the orgnal codng method n BOW, whch assgns descrptors to ust one codeword nearest to t. Although s smple, t yelds hgh quantzaton error. Then, soft assgnment (SA) [] s developed, wheren each descrptor s represented by all the codewords accordng to ther Eucldean dstances n Gaussan functon. To further mprove t, localzed soft assgnment codng (LSAC) [] was proposed by ntroducng localty constrant. Sparse codng (SC) [3] s another novel method to reduce quantzaton error, whch s realzed by reconstructng descrptors plus a sparse constrant to the codes. However, t s non-consstent and tme consumng. Then, localty-constraned lnear codng (LLC) [4] was proposed by ncorporatng localty constrant nto the obectve functon to encourage smlar descrptors have smlar codes. Whle Gao et al. [5] proposed Laplcan sparse codng (LSC) to preserve the consstence of codng. However, t s computatonally nfeasble. In order to meet the applcatons of large scale mage classfcaton, hgh dmensonal schemes have been proposed, such as Fsher kernel codng (FKC) [6] and super vector codng (SVC) [7]. It usually needs a large quantty of memory. Huang et al. [8] found that salency s one of the fundamental characterstcs of feature codng when combnng wth Max-poolng and then proposed salency-based codng (SaC), whch performs much better than classc assgnment schemes and more effcent than reconstructon based schemes. To mprove t, Wu et al. [9] further proposed group salent codng (GSC), wheren the latent structure nformaton of a codebook s explored by groupng neghborng codewords nto a group-code. Recently, a novel approach called local smlarty global codng (LSGC) [30] was proposed, whch uses the local smlartes between bases to obtan a nonlnear global smlarty measure between local descrptor and bases. From the above arguments, we can see that there are stll some lmtatons haven t been well solved n prevous works. We summarze t n Table. In ths paper, we propose a codng and poolng scheme wth low quantzaton, non-consstency, computatonal cost, ambguty, though ntroducng salency detecton and fuzzy reasonng rules, whch called fuzzy reasonng based salent codng (FRSC). In detal, salency detecton are used to generate salency maps whch are used to dvde mage nto salent and non-salent regon, and then combne two separated codebook clustered n them to produce a structured dctonary. Then, fuzzy reasonng rules are ntroduced to select the most salent and stable codewords to encode. By usng t, the underlyng manfold structure of descrptors can be well captured. Fnally, salency maps are used agan to locate the nterest obect, whch can be used to spatal poolng to ncorporate spatal nformaton. Thus, our new mproved BOW model can be obtaned by combng the above together. The remander of ths paper s organzed as follows. In Secton, we brefly analyze varous codng and poolng schemes. Secton 3 presents our codng and poolng approach. Then expermental results on the Caltech 0, Scenes 5, and UIUC Sport databases are provded n Secton 4. Fnally n Secton 5, conclusons are drawn, some future work and applcatons are dscussed. Related Work In ths secton, we brefly revew commonly used codng and poolng schemes. Let x (x R d ) be a d dmensonal descrptor, B d M = (b,b,,b M ) be a codebook wth M cluster centers, and u (u R d ) be the codng coeffcent vector of x, e.g., u be the response of x on codeword b. Hard assgnment codng: For a local descrptor x, only the closest codeword s used for codng, n whch Eucldean dstance s used., Ιf = argmn x b u, = =,, n () 0, others Soft assgnment codng: Each local descrptor s encoded by multple codewords usng the kernel functon of dstance between descrptors and codewords, such as Gaussan functon. exp( β x b ) u, = () m exp β x b k = ( k ) E-ISSN: Volume 3, 04

3 Table Comparson of prevous codng schemes. H: hgh, M: mddle, L: low. Quantzaton error Non-consstency Computatonal cost Ambguty HA[8] H L L H SA[] L L M H LSAC[] L L L H SC[3] L H H H LLC[4] L L L H LSC[5] L L H H FKC[6] L L H H SVC[7] L L H H SaC[8] H L L H GSC[9] L L L H LSGC[30] L L H H where β s the smoothng factor controllng the softness of the assgnment, and m [, n]. Localzed soft assgnment codng: It s an mproved verson of SA. Ther dfference s that SA encodes each descrptor across all the codewords whle LSAC confnes t to a local neghborhood around the coded descrptor. exp ( βdˆ ( x, b) ) u, = n exp ( ˆ βd( x, bl) ) (3) l= (, ), If Ν ( ) ˆ d x bl bl K x d( x, bl) =, others Sparse codng: It s a reconstructon based codng, whch use sparse constrant to allevate the quantzaton error. In detal, t represents a local descrptor by a lnear combnaton of a sparse set of bass vectors by solvng an l -norm regularzed approxmaton problem, whch can be solved by FS (Feature-sgn search algorthm) [3] and LD (Lagrange dual) [3]. = arg mn +λ u x Bu u (4) st.. u = where denotes the l -norm. Localty-constraned lnear codng: Further study [3] found that the localty constrant s more mportant than the sparse constrant. Thus, LLC was proposed by ntroducng the localty constrant, whch s obtaned by mnmzng the Eucldean dstance between each descrptor and codewords. u = arg mn x Bu + λ d u st.. u = (, ) dst x B d = exp σ (5) where dst(x, B) denotes the Eucldean dstance between x and b, σ s a parameter controllng the weghtng vector d. Furthermore, a smplfed and fast mplementaton was proposed to reduce the computaton cost. u = arg mn x Bu (6) st.. u = where B s K closest codewords to x. Salent codng: Its man dea s employng the dfference between the closest codeword and the other K- closest codewords to reflect salency. Thus, a local descrptor can be represented as: x b Ψ ( x, b ) = K x b k K k (7) Ψ ( xb, ),Ιf = arg mn ( x bl ) u ll, n, = 0, others [ b where, b,, b k ] s the K closest codewords to x. Group salency codng: Hard assgnment used n SaC s coarse for feature codng. Thus, group codng was ntroduced n GSC, whose man dea s calculatng the salency response of a group of codewords, and the response s fed back to all the codewords n the group, fnally, the maxmum of all responses are calculated accordng to dfferent group szes. k k Φ ( x), If b g( x, K) u, = 0, others (8) G+ k k Φ x = x b k+ t x b k ( ) ( ) t= E-ISSN: Volume 3, 04

4 Salency detecton Feature extracton on salent and nonsalent regons respectvely Classfcaton by lnear SVM Salency detecton based spatal poolng Fuzzy reasonng based salent codng Clusterng codebook on salent and nonsalent features respectvely Fg. The flowchart of the proposed method. where g(x, K) denotes the K closest codewords of x, G s the maxmum group sze. A poolng operaton s often needed to obtan an mage-level representaton when the codng responses of all local descrptors are calculated. Sum poolng: Average poolng: p p q = u (9) = q q =, = u (0) where q s the total number of local descrptors n an mage. Max poolng: p, = max u () The max poolng often gets better performance than sum and average poolng, such as n SC, LLC, SaC, GSC,LSAC. However, ts mechansm has not been fully studed n the lterature. 3 The Proposed Method The man components of the proposed approach s composed of three parts: salency detecton based structured codebook generaton, fuzzy reasonng based salent codng, and salency detecton based spatal poolng. The overvew of the proposed method s shown n Fg.. The detals of these three aspects are presented as follows. 3. Salency detecton based codebook generaton As we know, mages are usually corrupted by nose and there are often more than one obect n an mage wth dfferent shapes and occlusons, even n the same class. Thus, researchers dvde an mage nto, two tems whch called the correlated (or common) part and the specfc (or nosy) part respectvely. And the both parts are more robust and dscrmnatve for mage classfcaton because t captures complementary attrbutes n an mage. Inspred by these observatons, some low-rank based methods [3-33] are proposed. They use lowrank and spares technques to decompose local features of an mage or mages wthn each class nto a low-rank part and a sparse part, whch represent homogeneousness and dversty respectvely. However, they are tme consumng. In ths paper, we use salency map generated by salency detecton to decompose mages nto salent parts and non-salent parts. For computatonal effcency, we use effcent salency detecton method n [34]. Then, we extract SIFT descrptors n both parts and cluster by K-means to get two codebooks. Fnally, we combne them as a structured codebook to encode orgnal descrptors. Expermental results show that the structured dctonary has comparable representaton capablty wth low-rank based methods, whch are presented n Secton Fuzzy reasonng based salent codng Recently, salency based codng methods get satsfactory results due to ts effcency and stable representaton. However, they wll lose ther superorty n performance when codebook sze s relatvely large. Furthermore, they are also senstveness to unusual features, e.g., nosy features. Thus, we present a fuzzy reasonng based codng scheme to solve these problems n ths paper. In salency based codng, the response of a local descrptor s reflected by salency degree usng K closest codewords selected from the codebook. Then, only the maxmum response s preserved whle the low responses are suppressed n the later E-ISSN: Volume 3, 04

5 maxmum poolng operaton. Therefore, we thnk that only those largest responses are the meanngful responses. Salency s used to measure ths character n the orgnal salency based codng. However, f all the K closest codewords are near to the local descrptor, they have smlar salency value, t cannot reflect the salency n ths case, because all these K closest codewords are needed to represent the local descrptor. Thus, we ntroduce fuzzy reasonng rules to reflect t. Frst, we use d to denote the Eucldean dstance between the local descrptor and the th closest codeword, and s to denote the salency value of each local descrptor on the K closest codewords. d s = () K dk K k Take K=5 for example, we can defne sx fuzzy rules whch are descrbed as follows: Rule-: If s s low, s s low, s 3 s low, s 4 s low, and s 5 s low, then none of the K closest codewords can represent the local descrptor ndependently. Rule-: If s s hgh, s s low, s 3 s low, s 4 s low, and s 5 s low, then only the closest codeword can represent the local descrptor ndependently. Rule-3: If s s hgh, s s hgh, s 3 s low, s 4 s low, and s 5 s low, then the two closest codewords can represent the local descrptor stably. Rule-4: If s s hgh, s s hgh, s 3 s hgh, s 4 s low, and s 5 s low, then the three closest codewords can represent the local descrptor stably. Rule-5: If s s hgh, s s hgh, s 3 s hgh, s 4 s hgh, and s 5 s low, then the four closest codewords can represent the local descrptor stably. Rule-6: If s s hgh, s s hgh, s 3 s hgh, s 4 s hgh, and s 5 s hgh, then all the fve closest codewords can represent the local descrptor stably. Low and hgh are fuzzy membershp functons shown n Eq. 3 and Eq. 4. Both of them are trapezod shapes and llustrated n Fg. 3., s< a s b Low ( s) =, a s< b a b 0, s b 0, s< a s a Hgh ( s) =, a s< b b a, s b (3) (4) M e 0 Low(s) a b Hgh(s) Fg.3. The fuzzy membershp functons Low(s) and Hgh(s). Note that all the salency values are normalzed to n ths paper. And the two parameters a and b are usually set to 0. and 0.8 respectvely. Then, let the fuzzy truth value F be defned below: F = Low s Low s Low s Low s Low s F = Hgh s Low s Low s Low s Low s F = Hgh s Hgh s Low s Low s Low s F = Hgh s Hgh s Hgh s Low s Low s F = Hgh s Hgh s Hgh s Hgh s Low s F = Hgh s Hgh s Hgh s Hgh s Hgh s where product nference engne [35] s used to realze the fuzzy reasonng. After all the fuzzy truth values obtaned, we can determne whch codewords can be used to encode by the largest fuzzy truth value. Then, the prevous codng schemes can be used to encode, here, the SaC s adopted for effcency. Thus, our FRSC can be defned by: Case-: If max { F, F, F3, F4, F5, = F FRSC = { c,0,0,0,0 }; Case-: If max { F, F, F3, F4, F5, = F3 FRSC = { c, c,0,0,0 }; Case-3: If max { F, F, F3, F4, F5, = F4 FRSC = { c, c, c3,0,0 }; Case-4: If max { F, F, F3, F4, F5, = F5 FRSC = { c, c, c3, c4,0 }; Case-5: If max { F, F, F3, F4, F5, = F or max { F, F, F3, F4, F5, = F6 FRSC = { c, c, c3, c4, c5} ; c = Hgh s ( ) s E-ISSN: Volume 3, 04

6 Orgnal Image Salency Map Salent Regon Non-Salent Regon + = Fg.4. The dagram of our SSP. The colored shapes denote codewords. Fnally, max poolng s used to obtan the fnal codng responses for each local descrptor. 3.3 Salency detecton based spatal poolng Current state-of-the-art mage classfcaton systems are usually usng spatal pyramd matchng (SPM) to ncorporate the spatal nformaton, n whch pools low-level mage features over pre-defned coarse spatal bns, such as three levels of,, and 4 4. In ths paper, we propose a salency detecton based spatal poolng (SSP) approach for mage classfcaton. In contrast to SPM poolng, our SSP frst extracts the nterest obect n an mage by salency detecton n [34], then pools the codng responses obtaned n the prevous subsecton separately n the salent regon (nterest obect) and the non-salent regon (background) to form the mage-level representaton wth spatal nformaton, whch s shown n Fg.4. Obvously, our SSP tends to produce more consstent mage representaton than SPM poolng. 4 Expermental Result Ths experment ams to verfy that ) the structured dctonary can mprove the classfcaton performance; ) the proposed fuzzy reasonng based salent codng can produce comparable or even better performance than LLC, GSC, LSC, whch are wldly used or the state-of-the-art; ) the proposed SSP can perform better than SPM. We choose LLC, SaC, GSC, and LSAC for comparson. Note that all of them are effcent codng schemes. The followng three datasets are used for test: Caltech 0, Scenes 5, and UIUC Sport. Some example mages of these three datasets are shown n Fg.5. We frst study the E-ISSN: Volume 3, 04

7 Caltech 0: cup Caltech 0: wld cat Scenes 5: suburb Scenes 5: offce UIUC Sport: croquet UIUC Sport: salng Fg.5 Some example mages of the test datasets. E-ISSN: Volume 3, 04

8 Classfcaton accuracy [%] Orgnal Codebook Structured Codebook SaC GSC LLC (a) Classfcaton accuracy [%] SPM SSP SaC GSC LSAC Fg.6 Comparson between (a) orgnal codebook and structured codebook n dfferent codng schemes; (b) SPM and SSP under dfferent codng methods wth the orgnal codebook. (b) Classfcaton accuracy [%] LSAC LLC SaC GSC FRSC (a) Classfcaton accuracy [%] LSAC LLC SaC GSC FRSC Classfcaton accuracy [%] LSAC LLC SaC GSC FRSC (c) Fg.7 Performance comparson of varous codng strateges (a) under dfferent codebook sze wth orgnal codebook and SPM on the Caltech 0 dataset; (b) under dfferent codebook sze wth structured dctonary and SSP on the Scenes 5 dataset; (c) under dfferent codebook sze wth structured dctonary and SSP on the UIUC Sport dataset. (b) proposed method n the Caltech 0 dataset wth an n-depth analyss, ncludng dfferent codebook, codng and spatal poolng, and then combne them together n the other datasets. For far comparson, all the tested codng methods are mplemented n a unfed framework. Thus, the consstency of all the confguratons other than the codng part can be guaranteed. In our framework, SIFT descrptor s extracted from mages on a grd wth step sze of 6 pxels under 6 6 scale. Codebook s generated by standard K-means clusterng, wheren the subset of descrptors used for clusterng s randomly sampled from all descrptors n the database. Lb-lnear SVM s adopted for effcent classfcaton, wheren the penalty coeffcent s set to as most methods dd. As the other methods dd, we repeat the experment 0 tmes wth dfferent tranng and testng sets, then report the average accuracy and the standard devaton as the results. All the tests are conducted n a computer wth an Intel Core Duo.83 GHz CPU and GB RAM. 4. Results on the three datasets Caltech 0: Ths s wldly used dataset wth 9,44 mages n 0 classes ncludng a background class, whch contans anmals, vehcles, flowers, etc., and E-ISSN: Volume 3, 04

9 wth hgh ntra-class appearance shape varablty. Each category contans mages from 3 to 800. The average mage resoluton s pxels. In ths dataset, codebook sze s set as 04 when comparng dfferent part. Fg.6(a) shows the performance of dfferent codng schemes wth orgnal codebook and structured codebook respectvely. As shown, the structured codebook s slghtly better than the orgnal codebook. The classfcaton performance between SSP and SPM under dfferent codng methods wth orgnal codebook s shown n Fg.6(b), n whch we also fnd that our SSP s slghtly better than SPM. Then we show the results of varous codng strateges under dfferent codebook sze wth orgnal codebook and SPM n Fg.7(a). The proposed FRSC almost performs the same wth LSAC, but outperforms the other three codng schemes. We further make a comparson on the computaton cost, whch s shown n Table. We can see that the proposed FRSC s slghtly faster than LSAC. Table Computaton cost on Caltech 0 per mage. Method LLC SaC GSC LSAC FRSC ms Scenes 5: It s a natural scene dataset ncludng 4,485 mages fallen nto 5 scene categores, the number of mage per class ranges from 00 to 400. It contans bedroom, suburb, ndustral, ktchen, lvng room, coast, forest, hghway, nsde cty, mountan, open country, street, tall buldng, offce, and store. Followng the standard settng, 00 mages per class are used for tranng and the rest for testng. Combng our structured codebook, fuzzy reasonng codng, and SSP together, classfcaton accuracy under dfferent codebook sze s compared n Fg.7(b). As seen, our FRSC s slghtly better than LSAC when wth large (048) codebook sze, and outperforms the others. UIUC Sport: It s a sport event dataset, whch contans 8 categores ncludng badmnton, bocce, croquet, polo, rock clmbng, rowng, salng, and snowboardng. It contans 79 mages and the sze of each class vares from 37 to 50. The mage resoluton s hgher than the above two datasets. We randomly choose 70 mages for tranng and the remander for testng. Comparson results are shown n Fg.7(c). Agan, the proposed FRSC gets best performance, although t stll obtans smlar result wth LSAC. 4. Applcaton n elevator vdeos Vdeo event classfcaton s also an mportant computer vson problem. In ths paper, we further extend our FRSC nto event classfcaton n elevator vdeo. We frst select some vdeos from elevator, ncludng empty, full loadng, volence. Some example vdeo frames are shown n Fg.8. Each vdeo event contans 0 frames. In our experment, half of the frames n each vdeo are used for tranng and the other half for testng. SIFT descrptors are extracted for each frame on a dense grd, every 4 pxels and for 5 scale levels. To ncorporate spatal nformaton, the lnear verson of SPM kernel wth three levels of,, and 4 4 s adopted. Codebook sze s set to 56, whch s produced by k-means. Fnally, Lb-lnear SVM s adopted for event classfcaton. The classfcaton results are shown n Table 3. As seen, our FRSC based method acheves good performance n such smple vdeo events, whch shows potental applcaton n ntellgent elevator vdeo survellance, such as overload detecton, volence detecton, vdeo summarzaton. (a) Empty (b) Full loadng (c) Volence Fg.8 Some example vdeo frames. Table 3 Classfcaton accuracy of elevator vdeo event classfcaton. Event Accuracy Empty 00% Full loadng 99% Volence 98.6% 5 Concluson To allevate the ambguty and non-robustness problem n salency based codng, an mproved salent codng named FRSC was proposed n ths E-ISSN: Volume 3, 04

10 paper. Frst, we ntroduce effcent salency detecton method and use the generated salency maps to measure the salency degree of each mage, then dvde each mage nto salent and non-salent regon to get a structured dctonary. Then, FRSC was proposed to mprove the stablty by ntroducng fuzzy reasonng rules. Fnally, a salency detecton based spatal poolng scheme was proposed to ncorporate spatal nformaton to obtan a more compact mage representaton. Experment on several common used datasets demonstrated the effectveness of the proposed codng approach. At the same tme, our method s more effcent than the reconstructon based codng schemes. We further apply t nto elevator vdeo event classfcaton and acheved good performance, whch demonstrate the potental applcaton n ntellgent elevator vdeo survellance, such as overload detecton, volence detecton, vdeo summarzaton. Our future works wll focus on conductng extensve experment on more complcated elevator vdeo events classfcaton. Acknowledgement The authors would lke to express ther sncere thanks to the anonymous revewers for ther nvaluable suggestons and comments to mprove ths paper. Ths work s supported by scence and technology plannng proect of chongqng bureau of qualty and technology supervson. References: [] Z. Wu, Q. Ke, M. Isard, J. Sun, Bundlng features for large scale partal-duplcate web mage search, n: IEEE Conference on Computer Vson and Pattern Recognton, 009, pp.5 3. [] Jang Yu-Gang, Ngo Chong-Wah, Yang Jun, Towards optmal bag-of-features for obect categorzaton and semantc vdeo retreval, n: Proceedngs of the 6th ACM Internatonal Conference on Image and Vdeo Retreval, 007, pp [3] R. Collns, A. Lpton, T. Kanade, H. Fuuyosh, D. Duggns, Y. Tsn, D. Tollver, N. Enomoto, O. Hasegawa, A system for vdeo survellance and montorng, techncal report: CMU-RI-TR- 00-, Pttsburgh, PA, 000. [4] V. I. Pavlovc, R. Sharma, T. S. Huang, Vsual nterpretaton of hand gestures for humancomputer nteracton: A revew, IEEE Transactons on Pattern Analyss and Machne Intellgence, vol. 9, 997, pp [5] R. Kosala, H. Blockeel, Web mnng research: A survey, ACM SIGKDD Exploratons Newsletter, vol., no., 000, pp. 5. [6] A. K. Jan, A. Ross, S. Prabhakar, An ntroducton to bometrc recognton, IEEE Transactons on Crcuts and Systems for Vdeo Technology, vol. 4, no., 004, pp.4 0. [7] G. Csurka, C. Bray, C. Dance, L. Fan, Vsual categorzaton wth bags of keyponts, n: European Conference on Computer Vson, 004. [8] S. Lazebnk, C. Schmd, J. Ponce, Beyond Bags of Features: Spatal Pyramd Matchng for Recognzng Natural Scene Categores, n: IEEE Conference on Computer Vson and Pattern Recognton, 006. [9] /Caltech0/. [0] lazebnk/research/scenecategores.zp/, 006. [] Datasets/Caltech56/. [] [3] C. Harrs, M. Stephens, A combned corner and edge detector, n: Proceedngs of the Fourth Alvey Vson Conference, 988, pp [4] K. Mkolaczyk, C. Schmd, Scale and affne nvarant nterest pont detectors, Internatonal Journal of Computer Vson, vol. 60, no., 004, pp [5] D. G. Lowe, Dstnctve Image Features from Scale-Invarant Keyponts, Internatonal Journal of Computer Vson, vol. 60, no., 004, pp.9-0. [6] B. T. Navneet Dalal, Hstograms of Orented Gradents for Human Detecton, n: IEEE Conference on Computer Vson and Pattern Recognton, 005, pp [7] S. P. Lloyd, Least squares quantzaton n PCM, IEEE Transactons on Informaton Theory, vol. 8, no., 98, pp [8] C. Cortes, V. Vapnk, Support-vector network, Machne Learnng, 995, pp [9] [0] Y. Huang, Z. Wu, L. Wang, T. Tan, Feature Codng n Image Classfcaton: A Comprehensve Study, IEEE Transactons on Pattern Analyss and Machne Intellgence, vol. 36, no. 3, 04, pp [] J. C. Gemert, J. Geusebroek, C. J. Veenman, A. W. M. Smeulders, Kernel Codebooks for Scene E-ISSN: Volume 3, 04

11 Categorzaton, n: European Conference on Computer Vson, 008, pp [] L. Lu, L. Wang, X. Lu, In Defense of Softassgnment Codng, n: Internatonal Conference on Computer Vson, 0, pp [3] J. Yang, K. Yu, Y. Gong, T. Huang, Lnear spatal pyramd matchng usng sparse codng for mage classfcaton, n: IEEE Conference on Computer Vson and Pattern Recognton, 009, pp [4] J. Wang, J. Yang, K. Yu, F. Lv, T. Huang, Y. Gong, Localty-constraned lnear codng for mage classfcaton, n: IEEE Conference on Computer Vson and Pattern Recognton, 00, pp [5] S. Gao, I. W. Tsang, L. Cha, Laplacan Sparse Codng, Hypergraph Laplacan Sparse Codng, and Applcatons, IEEE Transactons on Pattern Analyss and Machne Intellgence, vol.35, no., 03, pp [6] F. Perronnn, J. Sanchez, T. Mensnk, Improvng the fsher kernel for large-scale mage classfcaton, n: European Conference on Computer Vson, 00, pp [7] X. Zhou, K. Yu, T. Zhang, T. S. Huang, Image classfcaton usng super-vector codng of local mage descrptors, n: European Conference on Computer Vson, 00, pp [8] Y. Huang, K. Huang, Y. Yu, T. Tan, Salent codng for mage classfcaton, n: IEEE Conference on Computer Vson and Pattern Recognton, 0, pp [9] Z. Wu, Y. Huang, L. Wang, T. Tan, Group Encodng of Local Features n Image Classfcaton, n: Internatonal Conference on Pattern Recognton, 0, pp [30] A. Shaban, H. R. Rabee, M. Faratabar, M. Ghazvnnead, From Local Smlarty to Global Codng; An Applcaton to Image Classfcaton, n: IEEE Conference on Computer Vson and Pattern Recognton, 03, pp [3] H. Lee, B. Alexs, R. Raat, Ng. Andrew Y, Effcent sparse codng algorthms, n: Conference on Neural Informaton Processng Systems, 006, pp [3] C. Zhang, J. Lu, Q. Tan, C. Xu, H. Lu, S. Ma, Image Classfcaton by Non-Negatve Sparse Codng, Low-Rank and Sparse Decomposton, n: IEEE Conference on Computer Vson and Pattern Recognton, 0, pp [33] L. Zhang, C. Ma, Low-rank decomposton and Laplacan group sparse codng for mage classfcaton, Neurocomputng, 04. [34] S. Chen, W. Sh, W. Zhang, Vsual salency detecton va multple background estmaton and spatal dstrbuton, Optk, vol. 5, no., 04, pp [35] L.X. Wang, Adaptve Fuzzy Systems and Control: Desgn and Stablty Analyss, Prentce-Hall, Englewood Clffs, NJ, 994. [36] Reza Tavol, Classfcaton and Evaluaton of Document Image Retreval System, WSEAS Transactons on Computers, vol., no. 0, 0, pp [37] Mahmoud Al-Ayyoub, Duaa Alawad, Khaldun Al-Darabsah, Inad Alarrah, Automatc Detecton and Classfcaton of Bran Hemorrhages, WSEAS Transactons on Computers, vol., no. 0, 03, pp E-ISSN: Volume 3, 04