Evaluation, Enhancement, Development & Implementation of Content Based Image Retrieval Algorithms

Transcription

1 6. Image Retrieval 6.1 Intrductin The chapter deals with image features, characteristics f image databases used and prpsed methds fr image retrieval. The image retrieval has been carried ut n the basis f: Clr cdes f entire image Fregrund clr cdes Fregrund shape crrelatin Cmbinatin f fregrund clr cdes and shape crrelatin With selectable percentage prprtin f weight f fregrund clr cdes and fregrund shape crrelatin fr cmpsite similarity measure Similar face images cntaining cmplex backgrund The perfrmance f retrieval methds has been evaluated with Precisin, Recall, F measure and P R curves fr varius images f different classes and categries. Query respnses f sme example-images have als been presented. The methd and results f face extractin and similar-face-image retrieval are cvered lastly in the chapter. The retrieval f similar images needs t meet tw extreme requirements (i) Retrieve nly similar images and (ii) Retrieve as many as pssible similar images. The perfrmance indicatr fr the first requirement is Precisin whereas that f the secnd is Recall. (Sectin 3.3). An attempt t imprve Precisin by either incrprating stringent features r strict image similarity cmparisn-measures fr excluding dissimilar images ends up in exclusin f similar images as well, adversely affecting the Recall. On ther hand, an attempt t increase the number f retrieved similar images leading t imprvement in Recall by incrprating brader features r relaxed similarity-measures ends up in retrieving mre number f similar images alng with dissimilar images, adversely affecting Precisin. Thus a gd CBIR system shuld retain maximum pssible Precisin fr higher Recall fr large image databases cnsisting f variety f 108

2 images. Sectin Our Observatins and Sectin Our Appraches fr image retrieval may be referred fr details. The prpsed methds fr image retrieval are based n tw streams f techniques. The first stream fllws brader image descriptrs Clr cdes and crrespnding histgrams. Whereas, the secnd stream fllws reliable prcessing leading t precise detectin f prminent bundaries eventually revealing fregrund. Image retrieval based n whle image cmparisn has been carried ut with brader image descriptrs fr retrieving images having similar clr cde distributin. The ther methds are based n fregrund f the images. The extracted fregrund has been utilized fr cmparisn f bjects cntained in it. The crrelatin cefficients have been used as similarity cmparisn-measure fr matching shape f the fregrunds. The ther three methdlgies cmbine afresaid tw streams f techniques fr image retrieval. The first cmbinatinal technique is based n similarity cmparisn f fregrund clr cdes. The secnd cmbinatinal technique is based n clr-cdes and shape f the fregrund with selectable prprtin f weight f shape & clrcde in cmpsite similarity-measure. And the third ne is the cmbinatinal technique applied fr applicatin specific CBIR fr similar-face image retrieval. Thus, ur appraches fr image retrieval techniques explits reliable prcessing resulting precise features fr better Precisin and brader image features fr better Recall. The methds are nvel fr the extracted features and their use fr image retrieval. 6.2 Image Features Figure 47 shws the blck diagram fr extractin f features fr the develped CBIR system. Parameters given as input includes selected image name fr single image prcessing r flder name fr bulk prcessing and wavelet decmpsitin level. The extractin f image features can be carried ut either fr ne image r fr all images stred in a flder fr supprting additin f an image int existing image database and t facilitate bulk prcessing respectively. The clr cde features f image can be separately extracted. The ptin f extracting all features take ut all features - prminent bundaries based and clr cde based. The extracted features stred in apprpriate data structures are preserved in secndary strage as a file crrespnding t the image. The unknwn dimensin f varius extracted features (i.e. 109

3 number f cnturs, number f vertices in cnturs) and their prcessing enfrced exhaustive applicatins f prgramming-skills & utilizatin f Matlab-features. Input f Parameters Extractin f Clr Cde Features Extractin f All Features Strage f Features Single Image All Images f the Flder Figure 47. Blck diagram - Feature extractin. The extracted image features listed in Table 5, can be categrized as Image attributes Name (Path), type & dimensin. Clr features - Clr cdes, Histgrams Bundaries based features - Reginal attributes, edges Fregrund related attributes bundaries based and clr features fr fregrund Face regin features 110

4 Table 5. Extracted features. Sr. Features N. 1 Path f the image. 2 N f rws f the image. 3 N f clumn f the image. 4 Image type. 5 Nrmalized glbal histgram f the image. Fr R G and B channels. Nt utilized at present. 6 Nrmalized cumulative glbal histgram f the image. Fr R G and B channels. Nt utilized at present. 7 Clr cdes. 8 Nrmalized histgram f clr cdes. 9 Nrmalized cumulative Histgram f clr cdes. Nt utilized at present. 10 Thinned edges. 11 Fregrund regins f the image. 12 Backgrund regins f the image. 13 Watershed pixels. 14 Cmpsite prminence measure. 15 Labeled regins. 16 Labeled regins cnverted t RGB image. 17 Image categry. Presently describing face r unknwn. 18 Extracted face. 19 Regins crrespnding t clr cdes f whle image. 20 Srted histgram f clr cdes f whle image. 21 Regins indices, srted fr regin area. Regins are fund fllwing prminent bundary based apprach. 22 Reginal attributes fr regins f whle image. Area, Centrid, Bunding bx, Extrema, Extent, Slidity (given by Area/CnvexArea), Eccentricity, Cnvex hull, Minr axis length, Majr axis length, rientatin. 23 Nrmalized Un-segmented fregrund regin f the image. 24 Un-segmented fregrund f the image. 111

5 Table 5 (Cntd.). Extracted features. Sr. Features N. 25 Nrmalized glbal histgram f fregrund f the image. Fr R G and B channels. Nt utilized at present. 26 Nrmalized cumulative glbal histgram f fregrund f the image. Fr R G and B channels. Nt utilized at present. 27 Clr cdes fr fregrund. 28 Nrmalized histgram f clr cdes f fregrund. 29 Nrmalized cumulative Histgram f clr cdes f fregrund. Nt utilized at present. 30 Regins crrespnding t clr cdes f fregrund. 31 Srted histgram f clr cdes f fregrund. 32 Rati f tw axes f Extrema crrespnding t nrmalized face regin. Nt utilized at present. 33 Rati f ther tw axes f Extrema crrespnding t nrmalized face regin. Nt utilized at present. 34 Rati f Minr axis length t Majr axis length f the ellipse that has same nrmalized secnd central mments as the face regin. 35 Orientatin f face regin. Angle in degree between x-axis and the majr axis f the ellipse that has same secnd mments as the face regin. 36 Average f fregrund regin clr cde. 37 Rati f fregrund area t image area indicating percentage cntributin f fregrund regin in the image. 38 Tw additinal extra feature fields Incrprated fr future needs. The clr cdes are used t describe clr attribute f pixels f images. A clr cde represents a set f clrs f RGB clr space. Ttal f 27 clr cdes are used t represent entire range f RGB clr space. The pixels assigned with clr cdes effectively segments the image by frming regins cnsisting f pixels with same clr - cdes. Figure 48 and Figure 49 (a) illustrate the clr cdes assigned t pixels resulting int segmentatin f image. The labeled regins f identical clr (same clr cdes) are shwn with same clr in the segmented images. Figure 49 (b) Left is a 112

6 separated regin f clrs crrespnding t flwers whereas Figure 49 (b) Right is fr the regins crrespnding t green leaves. Refer Annexure 4, Sectin A-4.4 fr details f prpsed nvel clr cdes and results f segmentatin applied n images f standard databases f BSDB [Fwlkes, n line] [Martin, 2001] and SIMPLIcity [Wang, 2001]. Figure 48. Clr cde based segmentatin. a) b) Figure 49. Clr cde based segmentatin & regins crrespnding t tw clr-cdes. 6.3 SIMPLIcity Image Database [Wang, 2001] [SIMPLIcity, n line] - Classes & Characteristics The dataset cnsists f ttal 1000 images f 10 classes and 100 images per class. Images are f medium reslutins and reasnable size. The test set has used by many researchers fr the purpse f CBIR. Table 6 describes characteristics f image classes. 113

7 Table 6. SIMPLIcity Image Database [Wang, 2001] [SIMPLIcity, n line] - Classes and Characteristics Sr. N. Class Name Characteristics Sample Images 0 Tribal peple Single r grup f tribal-inhabitants in different pses with different backgrunds; Clred faces; typical tribal-dressing; 1 Seashre Sea, sand, sky with cluds; sea-shre bjects; Images cver distant bjects; 2 Sculpture Majrity f sculpture images captured as distant bjects having sky as backgrund; 3 Bus Mstly single bus, cvering majr prtin f images; Differently clred varius types f buses with different backgrunds. 4 Dinsaur Different types f dinsaurs with nntextured multi-clred backgrunds; 5 Elephant Single r multiple elephants in different backgrunds; 6 Flwer Different types and clred flwers cvering majr prtin f images; 7 Hrse Single r multiple hrses in different backgrunds; 8 Muntain Muntains and sky 9 Served Fd n Restaurant-table Typical images f different types f served fd n restaurant-tables; 114

8 6.4 ALOI Image Database [ALOI, n line] [Geusebrek, 2001] Amsterdam Library f Object Images (ALOI) prvides a cllectin f clr images f ne-thusand small bjects, recrded fr scientific purpses. Over ne hundred images per bject were captured systematically in cntrlled cnditins fr varied viewing angles, illuminatin angles and illuminatin clrs fr the sensry variatin in bject recrdings. The images were captured by ne f five lights turned n fr 15 different illuminatin angles, 12 different illuminatin clr cnfiguratin and 72 bject view pint variatins. Prduced smth variatins in intensity and shadws in different directins and parts f wide variety f bjects ffer a cmprehensive test set fr studying segmentatin and feature extractin issues. Figure 50 shws few sampleimages f an bject recrded with such variatins. The test set cnsisting f sme f the images f ALOI [ALOI, n line] [Geusebrek, 2001] database has been used as ne f the databases fr image retrieval techniques fr studying effects f afresaid variatins n fregrund shape and fregrund clr cdes. 6.5 Prpsed Techniques Figure 50. ALOI sample images [ALOI, n line] [Geusebrek, 2001]. The prpsed techniques, based n Clr cdes f entire image, Fregrund clr cdes, Fregrund shape crrelatin and Cmbinatin f fregrund clr cdes & shape crrelatin fllw steps mentined belw fr image retrieval. Like all CBIR techniques, incrprated image features and methd fr cmputatin f similarity measures differentiate methds and their respective perfrmances fr retrieval f images. The prpsed methds cnsist f three phases (i) Input reading (ii) methd specific image-feature reading & prcessing fr similarity measures and (iii) utput / presentatin. The methd-specific Step 3 and Step 4 are presented in respective sectins f prpsed methds. The generic steps fr prpsed methds are: 115

9 Step 1: Read - name f selected query image, name f selected target flder. and image t be searched fr. Perfrm validatins fr inputs. Step 2: Read names f files cntaining all image features. Step 3: Read (r extract) required image-features fr the query image. Step 4: Fr every image-feature-file f target flder, Read crrespnding image-features f the image-feature-file f target flder Calculate (dis)similarity_index i fr i th image f the database. Preserve path f data base image, needed fr display. Step 5: Read similarity cut-ff. Step 6: Srt calculated (dis)similarity indices in descending rder. Step 7: Cunt n f images having better similarity index than the similarity cutff. Prepare fr prper presentatin f results. Step 8: Display all similar images having better similarity index than the similarity cut-ff, in rder f decreasing similarity (frm left t right, rw wise). Algrithm 4. Generic steps fr prpsed image retrieval methds The GUI (Annexure 2) based develped CBIR system runs n a stand-alne machine. The database can be expanded by adding images int any flder which can be prcessed subsequently fr feature extractin at nce with a muse click. In absence f indexing mechanism, features are stred in files. A user will be prmpted t carry ut feature extractin f a query image, if nt dne earlier with help f the GUI fr that single image. The preprcessing f a query image is adpted by cnsidering the high time cmplexity f algrithms and repetitin in the experimentatins. Fr real time deplyment f the system, query preprcessing can be eliminated. Similarity cut-ff selectin is t be carried ut by user with GUI. Lwer similarity cut-ff signifies higher permitted dissimilarity in the retrieved images. The image-query respnse gets presented in a Matlab Figure-windw cntaining a grid f 4 x 4 thumbnail-images alng with their strage path. 6.6 Whle Image Clr Cdes Based CBIR The prpsed methd fr image retrieval enables user t perfrm search n the basis f clr attributes f entire image. The clr cde assigned t a pixel is designated fr bradly describing the clr f pixel. Being a brader descriptr, a clr cde can accmmdate pixel clr variatins withut affecting crrespnding clr 116

10 feature. The prpsed technique cmpares nrmalized glbal histgram f clr cdes cnstructed fr entire query image with that f images f the database fr measuring clr distributin similarity. The steps invlved in reading f image features, their prcessing and cmputatin f similarity index are shwn belw. These methd specific steps replace crrespnding generic Steps 3 & 4 f Algrithm 4, Sectin 6.5. The methd specific steps are: Step 3: Read (extract) whle image clr cde features f given query image. Step 4: Fr every image-feature-file f target flder, Read crrespnding whle image clr cde features f the imagefeature-file f target flder. Calculate (dis)similarity_index i = abs(hqj hij), fr 1 < j <= number f bins, Where, hqj indicates j th bin f nrmalized histgram f clr cdes fr the query image. hij indicates j th bin f nrmalized histgram f clr cdes fr i th image f database. Preserve path f data base image, needed fr display. Algrithm 5. Whle image clr cdes based image retrieval Perfrmance Evaluatin The perfrmance f the methd has been tested n image database f SIMPLIcity [Wang, 2001] [SIMPLIcity, n line] cnsisting f 1000 images. Exhaustive perfrmance evaluatin has been carried ut fr fur classes f database Bus, Hrse, flwer and dinsaur. Recall, Precisin and F measure are cmputed fr sample queries f each class f images. The perfrmance measures fr different similarity cut-ffs have been tabulated fr cmparisn. Average Recall, Average Precisin and Average F measures fr the class are tabulated and pltted alng with P R curves fr query respnses Class: Bus The perfrmance evaluatin n image database [Wang, 2001] [SIMPLIcity, n line] cnsisting f 1000 images has been shwn in Table 7 fr 11 varieties f query images [Wang, 2001] [SIMPLIcity, n line] fr different similarity cut-ffs & 117

11 56 queries The selected query images pssess variatins in bject-pses, number f fregrund bjects, bject-clrs, backgrunds and illuminatin cnditins. Table 7. Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Class - Bus. 300.jpg 310.jpg 358.jpg 315.jpg 319.jpg Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Prec isin p = rr / ttal F meas ure = 2 / (1/p + 1/r) jpg

12 Table 7 (Cntd.). Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Class - Bus. Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Preci sin p = rr / ttal F meas ure = 2 / (1/p + 1/r) 365.jpg jpg jpg jpg jpg The average Recall and average Precisin fr the query image class bus fr different similarity cut-ffs have been tabulated in Table 8. Crrespnding P R curves f sample queries f the table alng with average Precisin & average Recall are shwn in Figure 51. Average Precisin, average Recall and average F-measures fr different similarity cut-ffs fr the query images f Table 8 has been pltted in Figure

13 Table 8. Average Recall, Average Precisin & Average F -measure. (Whle image clr cdes). Class Bus. Similarity cut-ff Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r) P - R Curves ( Whle Image Clr Cdes) Buses Precisin , , Recall 300.jpg 310.jpg 358.jpg 315.jpg 319.jpg 326.jpg 365.jpg 388.jpg 366.jpg 344.jpg 369.jpg Average Figure 51. P- R curves (whle image clr cdes). Class - Bus. Average Precisin, Recall & F Measure at Different Similarity Cut-ff Values: Whle Image Clr Cdes - Buses Similarity Cut-ff Average Precisin Average Recall Average F-Measure Figure 52. Average Precisin, Average Recall, Average F measures verses Similarity cut-ffs. (Whle image clr cdes). Class Bus. Fllwing pints are bserved: The nature f btained P R curves matches with the practical P R curves. Stricter similarity cut-ff increases the Precisin at the cst f Recall. 120

14 Despite vast variatins in bus clrs, backgrund, pses and illuminatin cnditins, high recall with gd precisin is achievable fr many sample queries. Fr many queries, Precisin f 1.0 is achieved. The Precisin and recall are pr fr query images having range / yellw clred buses as ther image-classes cntain images f similar clr distributin. Range f average perfrmance measures fr the class 100 % f average Precisin fr 1 % f average Recall 48 % f average Precisin fr 46 % f average Recall Giving 52 % f fall in average Precisin t raise average Recall by 45 % The value f average Recall at average Precisin f 0.5 is 0.41 a reasnably gd perfrmance measures. Increase in the user selected similarity cut-ff indicates higher threshld fr the similarity measures fr retrieving very similar images. As a result, average Recall falls dwn due t eliminatin f images having lesser similarity and average Precisin increases as retrieved images are very similar due t higher cut-ff resembling Precisin Recall behavir f any practical CBIR system Respnse Examples: Class Bus The query respnses f a bus image [Wang, 2001] [SIMPLIcity, n line] at tw different similarity cut-ffs have been shwn in Figure 53 and Figure jpg Image Database Size: 1000 Respnse at Similarity cut-ff 60 rr 7, ttal 7, Ttal 100, r 0.07, p 1.00 Figure 53. Query respnse f a bus image at similarity cut-ff

15 310.jpg Image Database Size: 1000 Respnse at Similarity cut-ff - 40 rr 30, ttal 33, Ttal 100, r 0.3, p 0.91 Figure 54. Query respnse f a bus image at similarity cut-ff

16 Class: Hrse The perfrmance evaluatin n image database [Wang, 2001] [SIMPLIcity, n line] cnsisting f 1000 images has been shwn in Table 9 fr 6 varieties f query images [Wang, 2001] [SIMPLIcity, n line] fr different similarity cut-ffs & 29 queries The selected query images pssess variatins in bject-pses, number f fregrund bjects, bject-clrs, backgrunds and illuminatin cnditins. The fregrund bjects cnstitute relatively lesser percentage-prtin f the image cmpared t image class bus. Relatively less variatins in the backgrund clr distributins amng images f the class. Table 9. Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Class - Hrse. Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Prec isin p = rr / ttal F measur e = 2 / (1/p + 1/r) 700.jpg jpg jpg

17 Table 9 (Cntd). Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Similarity cut-ff Retrieved relevant images - rr Class - Hrse. ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Prec isin p = rr / ttal F measur e = 2 / (1/p + 1/r) 757.jpg jpg jpg The average Recall and average Precisin fr the query image class hrse fr different similarity cut-ffs has been tabulated in Table 10. The P R curves fr sample queries f the table and crrespnding average Precisin & average Recall are shwn in Figure 55. Average Precisin, average Recall and average F-measures fr different similarity cut-ffs fr the query images f Table 10 have been pltted in Figure 56. Similarity cut-ff Table 10. Average Recall, Average Precisin & Average F - measure. (Whle image clr cdes). Class Hrse. Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r)

18 P - R Curves (Whle Image Clr Cdes) Hrses Precisin , , Recall 700.jpg 725.jpg 744.jpg 757.jpg 701.jpg 707.jpg Average Expn. (Average) Figure 55. P- R curves (whle image clr cdes). Class- Hrse. Average Precisin, Recall & F Measure at Different Similarity Cut-ff Values: Whle Image Clr Cdes - Hrses Average Precisin Average Recall Average F - Measure Similarity Cut-ff Figure 56. Average Precisin, Average Recall, Average F measures verses Similarity cut-ffs. (Whle image clr cdes). Class Hrse. 125

19 Fllwing pints are bserved: The nature f btained P R curves is clse t ideal P R curves fr sme f the query images. Fr the variatins in clrs & pses f fregrund bjects, backgrund and illuminatin cnditins, high recall with gd precisin is achievable fr many sample queries. Stricter similarity cut-ff increases the Precisin at the cst f Recall. Fr many queries, Precisin f 1.0 is achieved. Fr many queries, Recall greater than f 0.7 is achieved. The uniqueness f the backgrund clrs cmbined with fregrund clrs ends up with clr distributins nt cmmn in the images f ther classes giving very gd perfrmance measures. Range f average perfrmance measures fr the class 87 % f average Precisin fr 13 % f average Recall 76 % f average Precisin fr 55 % f average Recall Giving nly 11 % f fall in average Precisin t raise average Recall by 42%. The expnentially extended trend line is well abve Precisin = (0.5) line and des nt intersect average Recall till its maximum pssible value, implies mre than 50% f average Precisin fr all average Recall values indicating exceptinally gd perfrmance measures Respnse Examples: Class - Hrse The query respnse f a hrse image [Wang, 2001] [SIMPLIcity, n line] at similarity cut-ff 25 has been shwn in Figure 57. The Recall f 71% with Precisin f 73% at similarity cut-ff f 25 is remarkable, giving 0.72 as F-measure. The query respnse f anther image cnsisting f tw hrses has been shwn in Figure 58 giving 100 % precisin with Recall f 14%. 126

20 725.jpg Image Database Size: 1000 Respnse qt Similarity cut-ff - 25 rr 71, ttal 96, Ttal 100, r 0.71, p 0.73 Figure 57. Query respnse f a hrse image at similarity cut-ff

21 Figure 57 (Cntd.). Query respnse f a hrse image at similarity cut-ff

22 Figure 57 (Cntd.). Query respnse f a hrse image at similarity cut-ff

23 701.jpg Image Database Size: 1000 Respnse at Similarity cut-ff 70 rr 14, ttal 14, Ttal 100, r 0.14, p 1.0 Figure 58. Query respnse f anther hrse image at similarity cut-ff Class - Flwer The perfrmance evaluatin n image database [Wang, 2001] [SIMPLIcity, n line] cnsisting f 1000 images has been shwn in Table 11 fr 10 varieties f query images [Wang, 2001] [SIMPLIcity, n line] fr different similarity cut-ffs 44 queries The selected query images pssess variatins in bject-pses, number f fregrund bjects, bject-clrs, backgrunds and illuminatin cnditins. The fregrund bjects generally cnstitute significant prtin f the image. 130

24 Table 11. Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Class - Flwer. Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure = 2 / (1/p + 1/r) 602.jpg jpg jpg jpg jpg

25 Table 11 (Cntd.). Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Class - Flwer. 696.jpg Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure = 2 / (1/p + 1/r) jpg jpg jpg jpg

26 The average Recall and average Precisin fr the query image class flwer fr different similarity cut-ffs has been tabulated in Table 12. The P R curves fr sample queries f Table 11 and crrespnding average Precisin & average Recall are shwn in Figure 59. Average Precisin, average Recall and average F-measure fr different similarity cut-ffs fr the query images f Table 12 have been pltted in Figure 60. Table 12. Average Recall, Average Precisin & Average F - measure. (Whle image clr cdes). Class Flwer. Similarity cut-ff Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r) P - R Curves (Whle Image Clr Cdes) - Flwers Precisin , , jpg 606.jpg 644.jpg 655.jpg 656.jpg 675.jpg 696.jpg 682.jpg 618.jpg 621.jpg Average Expn. (Average) Recall Figure 59. P- R curves (whle image clr cdes). Class- Flwer. 133

27 Average Precisin, Recall & F Measure at Different Similarity Cut-ff Values: Whle Image Clr Cdes - Flwers Average Precisin Average Recall Average F Measure Similarity Cut-ff Figure 60. Average Precisin, Average Recall, Average F measures verses Similarity cut-ffs. (Whle image clr cdes). Class Flwer. Fllwing pints are bserved: The nature f btained P R curves is clse t ideal P R curve fr sme f the query images and similar t practical P R curves fr majrity f query images. Fr the variatins in clrs & pses f fregrund bjects, backgrund and illuminatin cnditins, high recall with gd precisin is achieved fr many sample queries. Stricter similarity cut-ff increases the Precisin at the cst f Recall. Fr many queries, Precisin f 1.0 is achieved. Pr Precisin and Recall values fr image 682.jpg are because f blurred (filtered) backgrund cnstituting majr prtin f the image. At lwer cut-ffs, images f clred human faces and served restaurant fd als gets retrieved because f similar visual cues (Figure 61). Range f average perfrmance measures fr the class 90 % f average Precisin fr average 7 % f Recall 76 % f average Precisin fr 44 % f average Recall Giving nly 14 % f fall in average Precisin t raise average Recall by 37%. The expnentially extended trend line is well abve Precisin = (0.5) line and nt intersecting till average Recall value f 1, implies average precisin abve 50% fr all average Recall values gd perfrmance measures. 134

28 Respnse Example: Class Flwer The query respnse f a flwer image [Wang, 2001] [SIMPLIcity, n line] having typical fregrund clrs is shwn fr similarity cut-ff f 40 in Figure 61. The Recall is calculated fr ttal f 57 red / pink clred flwer images f the database. 655.jpg Image Database Size: 1000 Respnse at Similarity cut-ff 40 rr 17, ttal 21, Ttal 57, r 0.298, p Figure 61. Query respnse f a flwer image at similarity cut-ff

29 Class Dinsaur The perfrmance evaluatin n image database [Wang, 2001] [SIMPLIcity, n line] cnsisting f 1000 images has been shwn in Table 13 fr 6 varieties f query images [Wang, 2001] [SIMPLIcity, n line] fr different similarity cut-ffs & 32 queries The selected query images pssess variatins in bject-pses, bject-clrs. The backgrunds are simple, multi-clr tned & nn-textured. The fregrund bjects cnstitute relatively lesser percentage-prtin f the image cmpared t image class bus. Table 13. Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Class Dinsaur. Similar ity cutff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the databas e - Ttal Recall r = rr / Ttal Precisi n p = rr / ttal F measure = 2 / ( 1/ p + 1 /r) 406.jpg jpg jpg

30 Table 13 (Cntd.). Precisin, Recall & F measure at different similarity cut-ffs. (Whle image clr cdes). Class Dinsaur. Similar ity cutff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the databas e - Ttal Recall r = rr / Ttal Precisi n p = rr / ttal F measure = 2 / ( 1/ p + 1 /r) 429.jpg jpg jpg Table 14 lists the average Recall and average Precisin fr the query image class dinsaur fr different similarity cut-ffs. The P R curves fr sample queries f the table and crrespnding average Precisin & average Recall are shwn in Figure 62. Average Precisin, average Recall and average F-measures fr different similarity cutffs fr the query images f the Table 14 have been pltted in Figure 63. Similarity cut-ff Table 14. Average Recall, Average Precisin & Average F - measure. (Whle image clr cdes). Class Dinsaur. Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r) 137

31 P - R Curves ( Clr Cdes) Dinsaurs 1.2 Precisin , , jpg 408.jpg 415.jpg 429.jpg 455.jpg 475.jpg Average Expn. (Average) Recall Figure 62.. P- R curves (whle image clr cdes). Class- Dinsaur. Average Precisin, Recall & F Measure at Different Similarity Cut-ff Values: Whle Image Clr Cdes - Dinsaurs Average Precisin Average Recall Average F Measure Similarity Cut-ff Figure 63. Average Precisin, Average Recall, Average F measures verses Similarity cut-ffs. (Whle image clr cdes). Class Dinsaur. Fllwing pints are bserved: The nature f btained P R curves is clse t ideal P R curve fr sme f the query images and similar t practical P R curves fr majrity f query images. Fr the variatins in clrs & pses f fregrund bjects high recall with gd precisin is achieved fr many sample queries. Stricter similarity cut-ff increases the Precisin at the cst f Recall. 138

32 Fr many queries, Precisin f 1.0 is achieved. Range f average perfrmance measures fr the class 100 % f average Precisin fr 5 % f average Recall 75 % f average Precisin fr 33 % f average Recall Giving 25 % f fall in Precisin t raise Recall by 28 %. The expnentially extended trend line intersects average Precisin = (0.5) line at average Recall at value 0.95 (apprx.) implies gd perfrmance measures Respnse Example: Class - Dinsaur The query respnse f a dinsaur image [Wang, 2001] [SIMPLIcity, n line] is shwn fr similarity cut-ff f 40 in Figure jpg Image Database Size: 1000 Respnse at Similarity cut-ff 40. rr 30, ttal 35, Ttal 100, r 0.3, p 0.85 Figure 64. Query respnse f a dinsaur image at similarity cut-ff

33 Figure 64 (Cntd.). Query respnse f a dinsaur image at similarity cut-ff Respnse Examples: Other Classes The query respnses shwn in Figure 65 & Figure 66 can well illustrate the issue f subjectivity invlved in the intentin f user and image cntent descriptin. Is user intending t retrieve blue skied images r images f blue sky with white cluds? Or, is he aiming t get seashre images r images cntaining water? The answer will determine the number f relevant images retrieved and hence the Precisin, Recall & finally the perfrmance f the system. The Figure 67 t Figure 70 shw the query respnses f images f ther classes f SIMPLICity [Wang, 2001] [SIMPLIcity, n line]. 140

34 102.jpg Image Database Size: 1000 Respnse at Similarity cut-ff 60. Figure 65. Query respnse f a sea-shre image at similarity cut-ff jpg Image Database Size: 1000 Respnse at Similarity cut-ff 60. Figure 66. Query respnse f anther sea-shre image at similarity cut-ff

35 248.jpg Image Database Size: 1000 Respnse: (Similarity cut-ff 60) Figure 67. Query respnse f a sculpture image at similarity cut-ff jpg Image Database Size: 1000 Respnse: (Similarity cut-ff 70) Figure 68. Query respnse f an elephant image at similarity cut-ff

36 503.jpg Image Database Size: 1000 Respnse: (Similarity cut-ff 60) Figure 69. Query respnse f anther elephant image at similarity cut-ff jpg Image Database Size: 1000 Respnse at Similarity cut-ff - 60 Figure 70. Query respnse f served fd image at similarity cut-ff

37 10.jpg Image Database Size: 1000 Respnse at Similarity cut-ff 70 Figure 71. Query respnse f an image f a tribal man with clr painted n face at similarity cut-ff 60. The average Recall, average Precisin and average F-measures fr all test queries have been tabulated in Table 15 and class wise P R curves and methd average P R curves are shwn in Figure 72. The average Recall, average Precisin and average F-measures fr all test queries fr the methd have been pltted in Figure 73. Table 15. Average Recall, Average Precisin & Average F - measure. (Whle image clr cdes). All queries fr the methd. Similarity cut-ff Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r) 144

38 Average P - R Curves fr image-classes (Whle Image Clr Cdes) & Average fr the Methd 1.2 Average Precisin (Methd) 0.06, , 0.69 Class average - Bus Class average - Hrse Class average - Dinsaur Class average - Flw er Average fr the methd Average Recall ( Methd) Expn. (Average fr the methd) Figure 72. P- R curves (whle image clr cdes). All queries fr the methd. Average Precisin, Recall & F Measure at Different Similarity Cut-ff Values fr the methd: Whle Image Clr Cdes , , , , , , Similarity Cut-ff Average Recall (Methd) Average Precisin (Methd) Average F-measure (methd) Figure 73. Average Precisin, Average Recall, Average F measures verses Similarity cut-ffs. (Whle image clr cdes). All queries fr the methd. 145

39 6.6.2 Discussin The perfrmance has been evaluated n 1000 images f standard data base [Wang, 2001] [SIMPLIcity, n line] cnsisting f 10 classes f images fr ttal 161 queries with different similarity cut-ffs fr 33 query images f 4 different classes. The methd is rbust t illuminatin, pse and view pint variatins as it is based n brader clr descriptrs. The feature extractin and retrieval methds require lesser cmputatins cmpared t bundary detectin based prpsed methds. The brader descriptrs are characterized t yield higher Recall. S are the clr cdes. The methd wrks well even n images f pr quality and lw reslutins. The ranking f nearly similar images is high. The methd is nt sensitive t image scale and rtatin. The Precisin measures btained fr majrity f the queries f all fur classes are significantly high with gd Recall. Range f average perfrmance measures fr all queries f all classes 96 % f average Precisin fr 6 % f average Recall 69 % f average Precisin fr 44 % f average Recall Giving 27 % f fall in average Precisin t raise average Recall by 38%. The expnentially extended trend line intersects average Precisin = (0.5) line at average Recall at value 0.90 (apprx.) implies gd perfrmance measures. The results with prpsed methd are better than many reprted in literature. 6.7 Fregrund Clr Cdes Based CBIR Fregrund based image retrieval enables user t search images n the basic f bjects cntained in the image. The exclusin f backgrund narrws dwn subjectivity induced diversity abut the image cntent. Precisely detected fregrund encmpassing prminent bundaries yielding fregrund regin attributes and clr cdes f the fregrund are used as cmbined features fr image retrieval. The nrmalized histgram cnstructed fr fregrund regin is cmpared with that f image f image database. The algrithm is applied n clr cdes f fregrund fr 146

40 measuring clr distributin similarity f fregrund regins f images under cnsideratins. Fllwing methd specific steps replace crrespnding generic Steps 3 & 4 f Algrithm 4, Sectin 6.5 : Step 3: Read (r extract) fregrund clr cde features f given query image. Step 4: Fr every image-feature-file f target flder, Read crrespnding fregrund clr cde features f the imagefeature-file f target flder. Calculate (dis)similarity_index i = abs(hqj hij), fr 1 < j <= number f Bins, Where, hqj indicates j th bin f nrmalized histgram f clr cdes fr the query image hij indicates j th bin f nrmalized histgram f clr cdes fr i th image f database Stre path f data base image, needed fr display. Algrithm 6. Fregrund clr cdes based image retrieval Perfrmance Evaluatin The perfrmance f the methd has been tested n image database f SIMPLIcity [Wang, 2001] [SIMPLIcity, n line] cnsisting f 371 images. Exhaustive perfrmance evaluatin has been carried ut fr tw classes f database Bus and flwer. Recall, Precisin and F measure are cmputed fr sample queries f each class f images fr different similarity cut-ffs. Average Recall, Average Precisin and Average F measures fr the class are tabulated t analyze perfrmance f the methd fr given class f images. P R curves fr query respnses alng with Average Precisin and Average Recall are pltted fr perfrmance analysis. Average Recall, Average Precisin and Average F measures are als pltted fr different similarity cutff Class: Bus The perfrmance evaluatin n image database [Wang, 2001] [SIMPLIcity, n line] cnsisting f 371 images has been shwn in Table 16 fr 11 varieties f query images [Wang, 2001] [SIMPLIcity, n line] fr different similarity cut-ffs & 57 queries 147

41 The set f selected query images is same as the set used fr image retrieval using whle image clr cdes. Table 16. Precisin, Recall & F measure at different similarity cut-ffs. (Fregrund clr cdes). Class - Bus. Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure= 2/( 1/ p + 1/r) 300.jpg jpg jpg jpg jpg

42 Table 16 (Cntd.). Precisin, Recall & F measure at different similarity cut-ffs. (Fregrund clr cdes). Class - Bus. Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure= 2/( 1/ p + 1/r) 326.jpg jpg jpg jpg jpg jpg

43 Similarity cut-ff Table 17. Average Recall, Average Precisin & Average F -measure. (Fregrund clr cdes). Class Bus. Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r) The average Precisin and average Recall fr class Bus have been tabulated in Table 17 and pltted in Figure 74 alng with P R curves fr individual bus image query respnses. The average Recall, average Precisin and average F-measures with respect t similarity cut-ff fr all test queries f the class have been presented in Figure 75. P - R Curves ( Fregrund Clr Cdes) Buses Precisin , , jpg 310.jpg 358.jpg 315.jpg 319.jpg 326.jpg 365.jpg 388.jpg 366.jpg 344.jpg 369.jpg Average Expn. (Average) Recall Figure 74. P- R curves (Fregrund clr cdes). Class - Bus. 150

44 Average Precisin, Recall & F Measure at Different Similarity Cut-ff Values : Fregrund Clr Cdes - Buses Similarity Cut-ff Average Precisin Average Recall Average F Measure Figure 75. P- R curves (Fregrund clr cdes). Class - Bus. Fllwing pints are bserved: The nature f btained P R curves matches with the practical P R curves. Stricter similarity cut-ff increases the Precisin at the cst f Recall. Despite vast variatins in bus clrs, pses and illuminatin cnditins, high recall with gd precisin is achievable fr many sample queries. Fr all but ne queries, Precisin f 1.0 is achieved. The Precisin and recall measures have been imprved fr all but ne (319.jpg) query images. Imprvement in precisin is cntributed by tw factrs nly fregrund regin based cmparisn and reduced image database size. Range f average perfrmance measures fr the class 100 % f average Precisin fr 3 % f average Recall 75 % f average Precisin fr 40 % f average Recall Giving 25 % f fall in average Precisin t raise average Recall by 37 % The expnentially extended trend line intersects average Precisin = (0.5) line at average Recall at value 0.9 (apprx.) implies quite gd perfrmance measures fr images f the class Query Respnse Example: Class - Bus The query respnse f a bus image [Wang, 2001] [SIMPLIcity, n line] at similarity cut-ff f 25 is shwn in Figure 76. The Recall f 59 % with 97 % f Precisin is t be nted. 151

45 Query Image Image Database Size: 371 Respnse at Similarity cut-ff 25 rr 59, ttal 61, Ttal 100, r 0.59, p 0.97 Figure 76. Query respnse f a bus image at similarity cut-ff 25. (FGCC) 152

46 Figure 76 (Cntd.). Query respnse f a bus image at similarity cut-ff 25. (FGCC) Class Flwer The perfrmance evaluatin n image database [Wang, 2001] [SIMPLIcity, n line] cnsisting f 371 images has been shwn in Table 18 fr 10 varieties f query images [Wang, 2001] [SIMPLIcity, n line] fr different similarity cut-ffs & 51 queries The set f selected query images cnsists f 8 images used in the set fr image retrieval using whle image clr cdes. Tw images used in the first set have been replaced because f their inferir fregrund extractin. 153

47 Table 18. Precisin, Recall & F measure at different similarity cut-ffs. (Fregrund clr cdes). Class - Flwer. 606.jpg Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure= 2/( 1/ p + 1/r) jpg jpg jpg jpg jpg

48 Table 18 (Cntd.). Precisin, Recall & F measure at different similarity cut-ffs. (Fregrund clr cdes). Class - Flwer. Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure= 2/( 1/ p + 1/r) 682.jpg jpg jpg jpg Table 19. Average Recall, Average Precisin & Average F - measure. (Fregrund clr cdes). Class Flwer. Similarity cut-ff Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r)

49 The average Precisin and average Recall fr class Flwer have been tabulated in Table 19 and pltted in Figure 77 alng with P R curves fr individual query respnses. The average Recall, average Precisin and average F-measures fr all test queries f the class have been presented in Figure 78. P - R Curves (Fregrund Clr Cdes) Flwers Precisin , , Recall 606.jpg 644.jpg 655.jpg 656.jpg 696.jpg 675.jpg 682.jpg 618.jpg 621.jpg 640.jpg Average Expn. (Average) Figure 77. P- R curves (Fregrund clr cdes). Class- Flwer. Average Precisin, Recall & F Measure at Different Similarity Cut-ff Values : Fregrund Clr Cdes - Flwers Similarity Cut-ff Average Precisin Average Recall Average F Measure Figure 78. Avg. Precisin, Avg. Recall, Avg. F measures verses Similarity cut-ffs. (Fregrund clr cdes). Class Flwer. 156

50 Fllwing pints are bserved: The nature f btained P R curves matches with the practical P R curves. Stricter similarity cut-ff increases the Precisin at the cst f Recall. Despite vast variatins in fregrund clrs and illuminatin cnditins, high recall with gd precisin is achievable fr many sample queries. Fr many queries, Precisin f 1.0 is achieved fr reasnable Recall. Range f average perfrmance measures fr the class 100 % f average Precisin fr 16 % f average Recall 68 % f average Precisin fr 45% f average Recall Giving 32 % f fall in average Precisin t raise average Recall by 29 % The expnentially extended trend line intersects average Precisin = (0.5) line at average Recall at value 0.67 (apprx.) implies gd perfrmance measures fr images f the class Query Respnse Example: Class - Flwer The query respnses f tw different flwer images [Wang, 2001] [SIMPLIcity, n line] at lwest similarity cut-ff f 25 are shwn in Figure 79 & Figure 80 respectively. The selected query images are f typical fregrund clrs. 689.jpg Image Database Size: 371 Respnse at Similarity cut-ff 25 Figure 79. Query respnse f a flwer image at similarity cut-ff 25. (FGCC) 157

51 640.jpg Image Database Size: 371 Respnse at Similarity cut-ff 25 Figure 80. Query respnse f anther flwer image at similarity cut-ff 25. (FGCC) Discussin The perfrmance has been evaluated n 371 images f standard data base [Wang, 2001] [SIMPLIcity, n line] cnsisting f all images f tw classes and sme images frm ther classes fr ttal 115 queries with different similarity cut-ffs fr 21 query images f 2 different classes. The methd is nt suitable fr images cntaining very small fregrund bjects and bjects tuching t image bundaries which may nt be encmpassed by prminent bundaries. The extracted backgrund excludes backgrund and related features frm cmparisn enabling user t perfrm search based n bjects cntained in the image. The methd is rbust t illuminatin and less sensitive t pse and view pint variatins as it is based n brader clr descriptrs f the extracted fregrund. The feature extractin and retrieval methds require significant cmputatins. The perfrmance f the methd is nt sensitive t regins attached t fregrund bjects, because, fr given image, such regins can be made t cnstitute small percentage f ttal extracted fregrund by perfrming feature extractin at higher wavelet level. The lw reslutin and pr 158

52 quality f images affect fregrund extractin and hence perfrmance f the methd. The Precisin measures btained fr majrity f the queries are significantly high with gd Recall. The expnentially extended trend line intersects average Precisin = (0.5) line f bth classes at average Recall at values 0.9 & 0.67 (apprx.) respectively, imply gd perfrmance measures. 6.8 Fregrund Shape Crrelatin Based CBIR The nrmalized unsegmented fregrund regin has been utilized as the feature fr the image cmparisn. The methd specific steps, replacing Step 3 & Step 4 f algrithm specified in Algrithm 4, Sectin 6.5 are: Step 3: Read (r extract) nrmalized unsegmented fregrund regin features fr the query image. Call it Rq. Step 4: Fr every image-feature-file f target flder, Read nrmalized unsegmented fregrund regin features f the imagefeature-file f target flder. Let us call it RdI. Calculate crrelatin cefficients f Rq & RdI. Find the significant crrelatin cefficient. Calculate (dis)similarity_indexi = 100 abs(significant crrelatin cefficient f Rq & RdI) * 100 Algrithm 7. Fregrund shape crrelatin based image retrieval. Unsegmented fregrund regins have been btained by excluding backgrund regins fund in step 8 f Algrithm 3. The query respnse fr the methd is shwn in Figure 82. Nte that the perfrmance & results with 0 % weight f fregrund clr cde attributes in the cmpsite similarity cnstraint f next methd Fregrund Clr Cdes & Shape Crrelatin crrespnds t this methd f image retrieval Perfrmance Evaluatin The perfrmance f the methd has been tested n image database f SIMPLIcity [Wang, 2001] [SIMPLIcity, n line] cnsisting f 371 images. The Precisin, Recall and F measures are shwn in Table 19 fr 10 images f class Flwer and ttal f 20 queries. Crrespnding P R curves are pltted in Figure

53 Table 20. Precisin, Recall & F measure at different similarity cut-ffs. (Fregrund shape crrelatin). Class - Flwer. 606.jpg Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal 100 Recall r = rr / Ttal Precisin p = rr / ttal F measure= 2/( 1/ p + 1/r) jpg jpg jpg jpg jpg

54 Table 20 (Cntd.). Precisin, Recall & F measure at different similarity cut-ffs. (Fregrund shape crrelatin). Class - Flwer. 682.jpg Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal 100 Recall r = rr / Ttal Precisin p = rr / ttal F measure= 2/( 1/ p + 1/r) jpg jpg jpg Table 21. Average Recall, Average Precisin & Average F - measure. (Fregrund Shape crrelatin). Class Flwer. Similarity cut-ff Average Recall Average Precisin Average F measure = 2 / (1/Avg.p + 1/Avg.r) 161

55 P - R Curves ( Fregrund Shape Crrelatin) Flwer Precisin , , Recall 606.jpg 656.jpg 675.jpg 644.jpg 655.jpg 682.jpg 618.jpg 621.jpg 640.jpg 696.jpg Average Expn. (Average) Figure 81. P- R curves (Fregrund shape crrelatin). Class- Flwer. Fllwing pints are bserved: The P R curves btained are clse t ideal P R curves. Stricter similarity cut-ff increases the Precisin at the cst f Recall. Despite vast variatins in fregrund clrs and illuminatin cnditins, gd Recall with gd Precisin is achieved fr many sample queries. Fr all queries, Precisin f 1.0 is achieved fr reasnable Recall. Range f average perfrmance measures fr the class 100 % f average Precisin fr 9 % f average Recall 94 % f average Precisin fr 23% f average Recall Giving nly 6 % f fall in average Precisin t raise average Recall by 14 % The expnentially extended trend line is well abve average Precisin = (0.5) line implies gd perfrmance measures fr images f the class Query Respnse Example Figure 82 shws query respnse f a flwer image with similarity cut-ff 60 giving 34 flwer images based n shape cmparisn. It shuld be nted that fr same query image, whle image clr cde based apprach retrieves maximum f 4 flwer images whereas fregrund clr cde based apprach retrieves nly the query image even fr lwest similarity cut-ff. (Table 11 and Table 18 respectively.) 162

56 656jpg Image Database Size: 360 Respnse at Similarity cut-ff 60 Figure 82. Query respnse f a flwer image at similarity cut-ff 60. (FG shape crrelatin) 163

57 6.8.3 Discussin The methd is very sensitive t the shape f the fregrund. Fregrund bject shape altering regins affect the perfrmance f the methd very adversely. The methd gives very gd results fr images where fregrund is nt cntaining attached unwanted regins. E.g. images f ALOI database, images f class flwer and class bus f SIMPLIcity [Wang, 2001] [SIMPLIcity, n line] database. The methd may nt perfrm equally well fr the images like thse f BSDB [Fwlkes, n line] [Martin, 2001]. Relaxed similarity cut-ff ends up in pr Precisin. Recmmended similarity cut-ff is abve 60% fr better perfrmance. Shape matching being a stricter cnstraint, relatively higher Precisin and lwer Recall are bserved fr the methd. The shape crrelatin technique has been applied fr face regin matching fr the purpse f similar-face image retrieval. 6.9 Fregrund Clr Cdes & Fregrund Shape Based CBIR The prpsed methd cmpsitely explits tw fregrund features shape and clr cdes. The weight prprtin f these tw features in the similarity measures is selectable by the user. Thus, the cmpsite similarity measure signifies the prprtinate emphasis f user s search criteria. The nrmalized unsegmented fregrund regin and fregrund clr cdes have been utilized as the features fr the image cmparisn. The methd specific steps, replacing Step 3 & Step 4 f Algrithm 4, Sectin 6.5 are: Step 3: Read (r extract) fregrund clr cde features f given query image. Read (r extract) nrmalized unsegmented fregrund regin features fr the query image. Call it Rq. Step 4: Fr every image-feature-file f target flder, Read crrespnding fregrund clr cde features f the imagefeature-file f target flder. Calculate (dis)similarity_index i = abs(hqj hij), fr 1 < j <= number f bins, Where, hqj indicates j th bin f nrmalized histgram f clr cdes fr the query image 164

58 hij indicates j th bin f nrmalized histgram f clr cdes fr i th image f database Read nrmalized unsegmented fregrund regin features f the imagefeature-file f target flder. Let us call it RdI. Calculate crrelatin cefficients f Rq & RdI. Find the significant crrelatin cefficient. Calculate (dis)similarity_index1i = 100 abs (significant crrelatin cefficient f Rq & RdI) * 100 Read Fregrund_Clrcde_weight Calculate cmpsite (dis)similarity index as (dis)similarity_index I = (Fregrund_Clrcde_weight*(dis)similarity_index) + (( Fregrund_Clrcde_weight) *(dis)similarity_index1) Perfrmance Evaluatin Algrithm 8. Fregrund clr cdes & fregrund shape based image retrieval. The perfrmance f the methd fr different cmbinatins f weights f fregrund clr cde and fregrund shape crrelatin in cmpsite similarity index is tabulated fr an image (455.jpg [Wang, 2001] [SIMPLIcity, n line]) f Dinsaur class is shwn in Table 22. The respective P R curves are presented in Figure 83. Table 22. Precisin, Recall & F measure fr different prprtinate weights at different similarity cut-ffs. (Fregrund Clr cdes & fregrund shape crrelatin). Similarity cut-ff fr 455.jpg % Weight f FG CC in Similarity Index Retrieved relevant images - rr ttal retrieved images - ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure Ttal relevant images in the database, Ttal =

59 Table 22 (Cntd.). Precisin, Recall & F measure fr different prprtinate weights at different similarity cut-ffs. (Fregrund Clr cdes & fregrund shape crrelatin). Similarity cut-ff fr 455.jpg % Weight f FG CC in Similarity Index Retrieved relevant images - rr ttal retrieved images - ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure Ttal relevant images in the database, Ttal = P R Curves ( Dinsaur Image) : Cmpsite Similarity Criteria Fregrund Clr Cdes & Fregrund Crrelatin-Cefficients Precisin FG CC nly FG Shape nly Recall 0% FGCC & 100% FG Crr-Ceff 10 % FGCC & 90% FG Crr-Ceff 20% FGCC & 80% FG Crr Cefff 30% FGCC & 70% FG Crr-Ceff 40% FGCC & 60% FG Crr-Ceff 50% FGCC & 50% FG Crr-Ceff 80% FGCC & 20% FG Crr-Ceff 100% FGCC & 0% FG Crr-Ceff Figure 83. P- R curves fr different prprtinate weights f Fregrund clr cdes & fregrund shape crrelatin. 166

60 The suitable prprtin f weight fr best retrieval perfrmance is image specific. The user is given a chice f selecting the weight and altering the prprtin if required fr successive retrievals. The Precisin and Recall fr tw images f flwer given as queries with 70% - 30% & 30%-70% weight prprtin (Fregrund CC & Fregrund shape crrelatin) has been cmputed in Table 23. The high Precisin is ntewrthy. Table 23. Precisin, Recall & F measure fr tw different prprtinate weights at different similarity cut-ffs. (Fregrund Clr cdes & fregrund shape crrelatin). 606.jpg Simil arity cutff % Weight f FG CC in Similarity Index Retrieved relevant images - rr ttal retrieved images - ttal Recall r = rr / Ttal Precisin p = rr / ttal F measure = 2/( 1/ p + 1/r) Ttal relevant images in the database, Ttal = jpg Query Respnse Examples: Figure 84 and Figure 85 shw respective retrieval results fr a flwer image with 20 % and 10 % weight f fregrund clr cdes in the cmpsite similarity index. The reductin in weight f fregrund clr cdes and crrespnding increase in the weight f fregrund shape crrelatin results int retrieval f mre flwer images, nt necessarily while flwer images. Image retrieval with same query image fr fregrund clr cde methd (100 % weight f fregrund clr cdes), nly the query image gets retrieved (Table 11). The reductin f similarity cut-ff results int retrieval f mre images as shwn in Figure 86 & Figure

61 656jpg Image Database Size: 371 Respnse at Similarity Cut-ff 60 with Fregrund Clr Cde Weight 20 Figure 84. Query respnse f a flwer image at similarity cut-ff 60 with FG CC weight 20. (FGCC & FG shape crrelatin) 656jpg Image Database Size: 371 Respnse at Similarity Cut-ff 60 with Fregrund Clr Cde Weight 10 Figure 85. Query respnse f same flwer image at similarity cut-ff 60 with FG CC weight 10. (FGCC & FG shape crrelatin) 656jpg Image Database Size: 371 Respnse at Similarity Cut-ff 50 with Fregrund Clr Cde Weight 20 Figure 86. Query respnse f same flwer image at similarity cut-ff 50 with FG CC weight 20. (FGCC & FG shape crrelatin) 168

62 656jpg Image Database Size: 360 Respnse at Similarity Cut-ff 50 with Fregrund Clr Cde Weight 10 Figure 87. Query respnse f same flwer image at similarity cut-ff 50 with FG CC weight 10. (FGCC & FG shape crrelatin) 169

63 6.9.2 Discussin The methd explits advantages f all previusly prpsed methds. The fregrund based apprach eliminates unwanted, majr cntributing backgrund and related features enabling bject based search with additinal cnstraints f fregrund shape and clrs. The selectable prprtin f weight f fregrund clr cdes and fregrund shape ends up in gd perfrmance f the system Cmparisns - Query Respnses f Varius Algrithms The sectin prvides cmparisn f query respnses f varius prpsed methd f image retrieval fr same query images. The first example is fr cmparisn f respnse f varius methds fr a flwer image f SIMPLIcity [Wang, 2001] [SIMPLIcity, n line] image database whereas the secnd example is fr the ALOI image data base [ALOI, n line] [Geusebrek, 2001]. described in Sectin 6.4. Typical characteristics f the databases are Example 1 - SIMPLIcity image database [Wang, 2001] [SIMPLIcity, n line] Figure 88 t Figure 91 are the respective query respnses f prpsed fur methds fr same query-image with same similarity cut-ff f 60. The query respnse with whle image clr cdes based retrieval gives 2 similar flwer images shwn in Figure 88. The respnse f same query image with fregrund based clr cde apprach yields 15 similar red flwer images as shwn in Figure 89. The respnse with fregrund shape crrelatin methd is shwn in Figure 90 indicates retrieval f 39 flwer images (nt nly red) and 2 nn-flwer images. The fregrund clr cdes and fregrund shape based apprach with 30% weight t fregrund clr methd yields respnse shwn in Figure 91 giving 12 similarly shaped red flwer images. 170

64 606.jpg Image Database Size: 1000 Respnse - Whle Image clr cdes, Similarity cut-ff 60 Figure 88. Query respnse f a flwer image at similarity cut-ff 60. (Whle image clr cdes). 606.jpg Image Database Size: 371 Respnse - Fregrund clr cdes, Similarity cut-ff 60 Figure 89. Query respnse f same flwer image at similarity cut-ff 60. (Fregrund clr cdes). 171

65 606.jpg Image Database Size: 371 Respnse - Fregrund shape crrelatin, Similarity cut-ff 60 Figure 90. Query respnse f same flwer image at similarity cut-ff 60. (Fregrund shape crrelatin). 172

66 606.jpg Image Database Size: 371 Query respnse - Fregrund clr cdes & fregrund shape crrelatin, Similarity cut-ff 60 with 30% weight f fregrund clr cdes Figure 91. Query respnse f same flwer image at similarity cut-ff 60 with 30% weight f FGCC. (Fregrund clr cdes & fregrund shape crrelatin) Example 2 - ALOI image database [ALOI, n line] [Geusebrek, 2001] The effect f illuminatin changes and bject view pint variatins n image retrieval has been illustrated with fllwing query respnses f varius methds fr same query image a ty duck fr similarity cut-ff f 70. The majr prtin f the backgrund in the image causes pr Precisin fr whle image clr cde methd as shwn in Figure 92. The fregrund clr cde based apprach imprves the Precisin by giving yellw / white clred ty images as the respnse as shwn in Figure 93. The fregrund shape based methd fr retrieval gives gd Recall and Precisin with duck tys ranked higher, as shwn in Figure 94. The fregrund based cmpsite apprach with clr cdes and shape crrelatin with 30% weight f fregrund clr cdes give the best perfrmance as shwn in Figure

67 62_ l8c2.png Image Database Size: 112 Respnse - Whle Image clr cdes, Similarity cut-ff 70 Figure 92. Query respnse f an ALOI image at similarity cut-ff 70. (Whle image clr cdes). 174

68 Figure 92 (Cntd.). Query respnse f an ALOI image at similarity cut-ff 70. (Whle image clr cdes). 175

69 62_ l8c2.png Image Database Size: 112 Respnse - Fregrund clr cdes, Similarity cut-ff 70 Figure 93. Query respnse f same ALOI image at similarity cut-ff 70. (Fregrund clr cdes). 176

70 62_ l8c2.png Image Database Size: 112 Respnse - Fregrund shape crrelatin, Similarity cut-ff 70 Figure 94. Query respnse f same ALOI image at similarity cut-ff 70. (Fregrund shape crrelatin). 177

71 62_ l8c2.png Image Database Size: 112 Query respnse - Fregrund clr cdes & fregrund shape crrelatin, Similarity cut-ff 70 with 30% weight f fregrund clr cdes Figure 95. Query respnse f same ALOI image at similarity cut-ff 70 with 30% weight f FGCC. (Fregrund clr cdes & fregrund shape crrelatin) Discussin The whle image clr cde based apprach is well suitable fr finding images having similar fregrund-backgrund clr cde distributin. High Recall values are pssible because f brader clr descriptrs. The methd faces cnventinal limitatins because f nt cnsidering shapes r any ther reginal features. The fregrund regin based appraches enable fregrund bject based image search by cnsidering fregrund clr cdes and fregrund shape crrelatin. The fregrund clr cde based apprach permits retrieval f fregrund shape variant similar images. The fregrund shape crrelatin has been a stricter cmparisn and very sensitive t extracted fregrund shape. The cmpsite similarity measure f fregrund clr cdes & fregrund shape perfrms well fr majrity f query images. 178

72 6.11 Applicatin Specific CBIR - Similar Face Image Retrieval This sectin is an applicatin f prpsed methds fr applicatin-specific CBIR t retrieve similar face images frm images cntaining cmplex backgrund. The feature extractin phase cnsists f face regin extractin technique based n prminent bundaries detectin based fregrund separatin. The similarity measure fr face regins is the shape crrelatin cmparisn as described in Sectin 6.8. The high success rati f precise face regin extractin fr cmplex backgrunds and illuminatin variatins has been explited fr shape crrelatin based similarity cmparisn fr image retrieval Face Extractin frm Images Cntaining Cmplex Backgrund The sectin prpses nvel methd fr human frntal face extractin frm clr images characterized by uncntrlled illuminatin cnditins fr image capturing and cmplex backgrunds. The methd incrprates statinary Haar wavelet transfrm & prximity influence fr prminent bundaries detectin and watershed transfrm, prximity influence & mrphlgical peratins t separate fregrund / backgrund alng with regin and clr attributes fr human face extractin. The methd explits redundancy by calescing lcal clr cues f all clr channels t emphasize reliable prcessing t precisely detect the human face by aviding under-segmentatin and reducing ver-segmentatin & artifacts. The methd has been tested n face-image cllectin f standard database and n images captured by an amateur phtgrapher fr varius cmplex backgrunds having diversified textures, varied illuminatin cnditins and multiple backgrund bjects. The presented results shw the effectiveness f the methd fr frntal face extractin, prving it suitable as an input t applicatins like digital album catalgue, cntent based image retrieval, face recgnitin and facial expressin recgnitin. First well-thught ut algrithm fr image segmentatin wuld be a watershed algrithm [Beucher, 1979]. The watershed algrithm has inherent characteristic f finding lcal minima catchments basins prducing ver-segmentatin f regins and intrducing artifacts. Hence, watershed transfrm cannt be applied directly n images having textures, texture r smth clr tne variatins e.g. natural images. Any filtering-preprcessing befre applying watershed algrithm results int lss f infrmatin intrducing either leaks in the bundaries r spurius bundaries leading t imprper segmentatin. Hence, ther techniques required t be cmbined with 179

73 watershed algrithm t vercme shrt-cmings f watershed algrithm expliting advantages f the same. S has been dne in the prpsed methd. The face extractin is a prcess f islating face regin frm all ther regins. The perfrmance f the face extractin is challenged by characteristics like illuminatin variatins, shadws, skin-clred ther regins, face-shaped ther regins, multiple bjects, diversified indr & utdr backgrund textures, wide range f face skin clrs and hair clrs, different hair styles, different face rientatins, different image reslutins alng with afresaid image segmentatin issues as image segmentatin being the inevitable first step f the prcess. The issue f extractin f human face frm images captured in uncntrlled illuminatin cnditins having cmplex & nn-unifrm backgrund has been addressed in the prpsed methd by perfrming prper segmentatin fllwed by fregrund / backgrund separatin and face regin extractin. The image characteristics - illuminatin variatins and diversity & variatins in the backgrund textures impse challenges at the segmentatin and face extractin phase. The prpsed methd enfrces reliable prcessing f lcal clr cues f all clr and gray channels fr frming cntinuity preserving prminent bundaries incrprating Statinary Haar wavelet at varius levels. The prminent bundaries, prximity influence and watershed transfrms are cmpsitely used fr revealing fregrund frm the image. This fregrund may cnsist f human face, hair and attached regins due t cmplex & nn-unifrm backgrund. Varius regin attributes alng with clr attributes are used t extract the face. The methd vercmes issues f under-segmentatin by precise prcessing f lw level cues generating well lcalized, delineated leak-free bundaries which are further categrized as prminent bundaries encmpassing visually prminent regins in the image. The methd minimizes ver-segmentatin & artifacts prducing prper segmentatin needed fr face extractin in majrity f the cases cntaining illuminatin variatins. The incrpratin f Statinary Haar decmpsitin at varius levels makes methd suitable fr hierarchical framewrk. The perfrmance f the methd has been evaluated n face images f standard dataset Caltech 101 [Caltech, n line] [Fei-Fei, 2004]. 180

74 The Methd The prpsed methd explits redundancy by finally calescing lcal cues f R, G, B and Gray clr channels fr prminent bundaries detectin and fr frmatin f cmpsite watershed regins utilized fr face extractin. The regin attributes cnsidered fr determining the face regin are defined in the methd reginprps f Matlab 14 as fllws: Orientatin - The angle ( in degree) between the x-axis and the majr axis f the ellipse that has the same secnd-mments as the regin. Extent The prtin f the pixels in the bunding bx that are als in the regin. Eccentricity Measured fr the ellipse that has the same secnd-mments as the regin. It is the rati f the distance the fci f the ellipse and its majr axis length. The steps invlved fr the face extractin are: Step 1: Apply fregrund extractin Step 1 t Step 10 f Algrithm 3, Sectin 5.2. Step 2:Fr all regins f fregrund, Exclude small regins. Fr remaining regins, Mark a regin as face regin if rientatin > 70, Eccentricity > 0.3, extent < 0.95, axis _rati (Minr axis length / Majr axis length) > 0.4 and if regin cntains skin clr. Mark image categry as face. Step 3: Map face regin n the image t extract face image. Algrithm 9. Face extractin frm images cntaining cmplex backgrund. The threshlds are empirically determined t reflect shape attributes f face Results The methd has been tested n test set cnsists f 115 face-images f Caltech 101 face dataset [Caltech, n line] [Fei-Fei, 2004] cmprising f faces f 15 persns and ther high reslutin images captured by an amateur phtgrapher. The test set images are selected t cver varius illuminatin variatins and backgrunds. The images cntaining mustache and bearded face are excluded frm the test set. The Caltech dataset [Caltech, n line] [Fei-Fei, 2004] fund mst apprpriate as the data set because it cntains clr face images f medium size and reasnable reslutin. The multiple images f persns have been captured at varius indrutdr places with varius cmplex backgrunds under different illuminatin cnditins generating shadws and illuminatin variatins. Figure 96 illustrates results f varius phases f face extractin incrprating tw different levels f wavelet decmpsitin level 1 and level 2. The prminent bundaries detected fr Gray 181

75 channel has been shwn in Figure 96 (b). Varius regins f segmented clr channels have been shwn in Figure 96 (c) t Figure 96 (f). The cmpsite regins f segmented image are shwn in Figure 96 (g). The watershed pixels cnstituting regin islating bundaries have been shwn in Figure 96 (h). Figure 96 (i) and Figure 96 (j) crrespnds t separated backgrund and fregrund respectively. The fregrund is marked as black in the backgrund image and backgrund is marked as black in the fregrund image. The extracted faces are finally shwn in Figure 96 (k). In general, reductin f image dimensin by a large factr may adversely affect the segmentatin perfrmance due t interplated pixel clr values. The unaffected result f face extractin fr a size reduced high reslutin image has been shwn in Figure 97. Figure 98 illustrates unsuccessful face extractin due t imprper segmentatin because f extreme illuminatin variatins n the face. The extracted faces f sme f the images f test set have been shwn in Figure 99. The typically picked up images cntain distinctiveness like nn-face skin clr regins, ther face-shaped regins, skin-clred hair, ff-centered face, multiple backgrund bjects and mst imprtantly illuminatin variatins due t different lightning cnditins at indr-utdr lcatins. Table 24 depicts the effectiveness and rbustness f the methd fr successfully extracting faces fr 82.6 % f the images f the test set Discussin The prpsed methd is nvel fr prminent bundaries, prximity influence, Statinary Haar Wavelet, used fr generatin f cmpsite watershed regins fr fregrund separatin that is cmbined with face-regin & face-clr attributes fr frntal face extractin. The well lcalized and delineated cntinuity preserving prminent bundaries detected by precise and reliable prcessing f lw level clr cues f all clr and gray channels frm the basis f high (82.6 %) successful face extractin rati fr 115 test images f Caltech 101 [Caltech, n line] [Fei-Fei, 2004] dataset. The extractin f face has been tested n varius images cntaining perfrmance affecting characteristics like illuminatin variatins, shadws, skin-clred ther regins, face-shaped ther regins, multiple bjects, diversified indr & utdr cmplex-backgrund textures, wide range f 182

76 face skin clrs and hair clrs, different hair styles, different face rientatins etc. The statinary Haar wavelet decmpsitin at varius levels, prminent bundaries & prximity influence avids under-segmentatin and reduces ver-segmentatin & artifacts inherent characteristics f watershed transfrm. The methd results are nt affected due t interplatin peratin invlved in a size reductin f a high reslutin image by a large factr. As shape and clr f the face regin is largely altered due t mustache & bearded, the methd fails fr face extractin f such cases. Similarly, a dark face segmenting shadw als prduces ill results f face extractin. Still, precisely extracted faces with high perfrmance rati fr variety f images prves the suitability and versatility f the methd fr applicatins like digital album catalgue, cntent based image retrieval, face recgnitin and facial expressin recgnitin. (a) (b) (c) (d) Figure 96. Varius steps f face extractin. Left - statinary Haar wavelet decmpsitin at level 1. Right - statinary Haar wavelet decmpsitin at level 2. (a) Original image [Caltech, n line] [Fei-Fei, 2004]. (b) Detected prminent bundaries f gray channel. (c) Segmented regins f red clr channel. (d) Segmented regins f green clr channel. 183

77 (e) (f) (g) (h) (i) (j) (k) Figure 96 (Cntd.). Varius steps f face extractin. (e) Segmented regins f blue clr channel. (f) Segmented regins f gray clr channel. (g) Cmpsite segmented regins. (h) Crrespnding watershed pixels f (g). (i) Backgrund. (j) Fregrund. (k) Extracted face. 184

78 (a) (b) (c) (d) (e) Figure 97. Face extractin in high reslutin image reduced t 1/8 th f the riginal size. (a) Original image. (b) Watershed pixels. (c) Backgrund. (d) Fregrund. (e) Extracted face. (a) (b) (c) (d) Figure 98. Example f unsuccessful face extractin. (a) Original image [Caltech, n line] [Fei-Fei, 2004]. (b) Watershed pixels. (c) Backgrund. (d) Fregrund. (a) (b) (c) (d) (e) Figure 99. Face extractins f varius images with cmplex backgrund and nn-unifrm illuminatins. (a) Original images [Caltech, n line] [Fei-Fei, 2004]. (b) Watershed pixels. (c) Backgrunds. (d) Fregrunds. (e) Extracted face. 185

79 (a) (b) (c) (d) (e) Figure 99 (Cntd.). Face extractins f varius images with cmplex backgrund and nn-unifrm illuminatins. (a) Original images [Caltech, n line] [Fei-Fei, 2004]. (b) Watershed pixels. (c) Backgrunds. (d) Fregrunds. (e) Extracted face. 186

80 Table 24. Perfrmance evaluatin f face extractin methd fr varius images [Caltech, n line] [Fei-Fei, 2004]. Persn id Sample Images N. f Successful face extractin / Ttal images f the persn in the test set Successful face extractin % Persn 1 17 / Persn 2 7 / Persn 3 4 / Persn 4 11 / Persn 5 14 / Persn 6 5 / Persn 7 11 / Persn 8 4 / Persn 9 4 / Persn 10 3 / 5 60 Persn 11 4 / Persn 12 5 / Persn 13 4 / Other 2 / Ttal 95 / Similar Face Image Retrieval The methd described in Algrithm 7, Sectin 6.8 is applied n precisely extracted face regin fr cmparisn f shape crrelatin cefficients t retrieve similar face images. 187

81 Perfrmance Evaluatin The perfrmance measure Precisin & Recall have been tabulated in Table 25 fr 5 sample images f Caltech database. s Caltech [Caltech, n line] [Fei-Fei, 2004] Table 25. Precisin, Recall & F measure at different similarity cut-ffs. Similar face-image retrieval. Similarity cut-ff Retrieved relevant images - rr ttal retrieved images - ttal Ttal relevant images in the database - Ttal Recall r = rr / Ttal Precisin p = rr / ttal F - measure Image_0001.jpg Image_0007.jpg Image_0047.jpg Image_0079.jpg Image_0122.jpg

82 The average Precisin and average Recall at different similarity cut-ffs have been shwn in Table 26. The crrespnding P R curves are pltted in Figure 100. Table 26. Average Recall, average Precisin & average F measure at different similarity cut-ff. (Similar face image retrieval). Similarity cut-ff Average Recall Average Precisin Average F- measure P - R Curves (Similar Face Image Retreival) Precisin , , 0.23 Image_0001.jpg Image_0007.jpg Image_0047.jpg Image_0079.jpg Image_0122.jpg Average Recall Figure 100. P R curves. Similar face image retrieval. Fllwing pints are bserved: The nature f btained P R curves matches with the practical P R curves. Stricter similarity cut-ff increases the Precisin at the cst f Recall. Despite vast variatins in fregrund / backgrund clrs, pses and illuminatin cnditins, gd Recall with gd Precisin is achievable fr many sample queries. Fr all queries, Precisin f 1.0 is achieved at higher cut-ff. Range f average perfrmance measures fr all queries 100 % f average Precisin fr 19 % f average Recall 23 % f average Precisin fr 51 % f average Recall Giving 77 % f fall in average Precisin t raise average Recall by 32 % Lwer similarity cut-ffs less than 60, are nt recmmended fr gd perfrmance. 189

83 The expnentially extended trend line intersects average Precisin = (0.5) line at average Recall at value 0.45 (apprx.) implies gd perfrmance measures fr images f the class Query Respnse Example The query respnse fr retrieving similar-face images with 70 as similarity cut-ff is shwn in Figure 101. The illustrated respnse signifies Precisin f 75% fr 28% f Recall fr images having cmplex backgrunds and face-regins cnstituting a small prtin f images in a database f 115 images. Query Image Respnse at Similarity cut-ff 70 rr 6, ttal 8, Ttal 21, r 0.28, p Discussin Figure 101. Query Respnse similar face-image retrieval. The effectiveness f methds fr prminent bundaries detectin & fregrund separatin fr precisely extracting face by excluding cmplex backgrund has been utilized fr btaining face-features and similar-face images. The methd incrprates nly face shape-feature fr similarity cmparisn. Inclusin f ther face-features fr similarity cmparisn will imprve the perfrmance f the system Perfrmance Cmparisns with ther CBIR Techniques The relative perfrmance cmparisns f prpsed methds with the state f the art CBIR techniques is nt feasible because (i) unavailability and un-disclsure f cmprehensive technical details f state-f-the-art techniques which are cmmercial, prprietary r patented. (ii) Available n-line dems f sme f the CBIR systems 190