TELK OMNIKA V ol. 16, No . 3, J une 2018, pp . 1256 1263 ISSN: 1693-6930, accredited A b y DIKTI, Decree No: 58/DIKTI/K ep/2013 1256 W a velet-Based Color Histogram on Content-Based Ima g e Retrie v al Ale xander , Jeklin Hare fa* , Y ud y Purnama , and Har vianto Computer Science Depar tment, School of Computer Science , Bina Nusantar a Un iv ersity , J akar ta, Indonesia 11480 *Corresponding A uthor , email: jharef a@bin us .edu Abstract The g ro wth of image databases in man y domains , including f ash ion, biometr ic , g r aphic design, architecture , etc. has increased r apidly . Content Based Image Retr ie v al System (CBIR) is a technique used f or finding rele v ant images from those huge and unannotated image databases based on lo w-le v el f eatures of the quer y images . In this study , an attempt to emplo y 2 nd le v el W a v elet Based Color Histog r am (WBCH) on a CBIR system is proposed. Image database used in this study are tak en from W ang’ s image database containing 1000 color images . The e xper iment results sho w that 2 nd le v el WBCH giv es better precision (0.777) than the other methods , including 1 st le v el WBCH, Color Histog r am, Color Co-occurrence Matr ix, and W a v elet te xture f eature . It can be concluded that the 2 nd Le v el of WBCH ca n be applied to CBIR system. K e yw or ds: CBIR, W a v elet, Color Histog r am Cop yright c 2018 Univer sitas Ahmad Dahlan. All rights reser ved. 1. Intr oduction In this er a, the large n umber of digital images ha v e increased r apidly . This is because the large n umber of images data from v ar ious domains , such as f ashion, biometr ic , g r aphic design, architecture , etc. are in demand. One of the techniques f or digital image processing is Content Based Image Retr ie v al (CBIR). CBIR has been an activ e research area that helps to access and find the images from huge image database since 1990 [1]. The main idea of CBIR system is to e xtr act the lo w-le v el f eatures which are used to measure similar ity [2]. It applies the computer vision techniques in image retr ie v al based on lo w-le v el f eatures which can be automatically der iv ed from the f eatures presented in the images , such as color , te xture , or shapes [3]. The gener al systems in CBIR usually only use the lo w-le v el f eatures , such as color , te xture , and shape , and it doesn’t include an y semantic le v el. Color and T e xture are the tw o most common f eatures used in CBIR. The color histog r am is the first technique introduced in pix el domain [4]. It is commonly used in image compar ison because it is simple to compute and rob ust against small changes in camer a vie wpoint [5]. The te xtur e is also claimed to be the essential f eature in image retr ie v al because it can be decomposed into se v er al par ameters , such as coarseness , contr ast, and d irectionality [6]. Thus , man y researches ha v e used color and te xture f eatures in b uilding the CBIR system. Y ouness et al. [7] proposed a no v el method f or retr ie v al system using Gabor filters and 2-D ESPRIT method. In this study , each image is char acter iz ed b y the pair giv en using Gabor filters and the 2-D ESPRIT method applied to the or iginal image . This e xper iment achie v es a v er age precision of 80.19% using Brodatz images database . Ir ianto [8] used the Region Gro wing Segmentation f or searching and retr ie v e image from the database . Compared to Discrete Cosine T r ansf or m (DCT) images , this study can gain more efficient time and simplify the algor ithm. Lin et a l. [9] introduced three image f eatures which are: color , te xture , and color distr ib ution in order to de v elop a smar t retr ie v al system. This e xper iment calculates Diff erence betw een Pix els of Scan P atter n (DBPSP), Color Histog r am f or K-mean (CHKM) and Color Receiv ed December 23, 2016; Re vised F ebr uar y 5, 2018 ; Accepted F ebr uar y 23, 2018 DOI  10.12928/TELKOMNIKA.v16i3.7771 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 1257 Co-occurrence Matr ix (CCM) respectiv ely and enhance the perf or mance accur acy and simplified the image retr ie v al process . Ragupathi et al. [10] proposed a rob ust image retr ie v al system using the combination of diff erent f eature e xtr action methods , such as Color Histog r am (CH), Gabor T r ansf or m (GT), the combination of CH and GT , Cont our let T r ansf or m and the combination of CH and Contour let T r ansf or m. Hiremath and Pujar i [11] ha v e used the combination of color , te xture and shape f eatures within a m ultiresolution m ultig r id fr ame w or k. The research pro vides a rob ust f eature set and achiv e the highest precision compared to other retr ie v al systems . Another research comes from Manimala and Hemachandr an [12]. The y introduced the W a v elet Based Color Histog r am (WBCH) method in image retr ie v al which combines the HSV color and Gabor te xture f eatures of the image . The study giv es a promising result which pro v ed that WBCH has be tter an a v er age precision compared to the other fiv e methods (0.762). But this method only limited to the first le v el of WBCH. This paper attempts to impro ving the a v er age precision of the retr ie v al system b y changing the w a v elet le v el from the first le v el to the second le v el and third le v el of the w a v elet in order to obtain more precision. 2. Resear c h Method 2.1. Materials Data set used in this study is W ang’ s image database which is also one of the standard databases f or CBIR that contains 1000 images from the Corel image database represented with RGB color space . The images w ere divided into 10 categor ies which are Afr ican P eople , Beach, Buildings , Buses , Dinosaurs , Elephants , Flo w ers , Horses , Mountains , and F ood with JPEG f or mat and usually used in a gener al pur pose image database f or e xper imentation. 2.2. Methods Basically , there are tw o steps f or compar ing each image in the database and quer y image , which are: F eature Extr action and Similar ity Matching. F or the f eature e xtr action step , it is used to e xtr act the images f eatures f or classifying the objects . Similar ity Matching is used to get a result that is visually similar [13]. F eature that used in this study are color and te xture , while f or similar ity matching using Histog r am Intersection. Based on the Figure 1, t he proposed method will be applied to each database images and the quer y images . Firstly , e v er y f eature in each image will be e xtr acted first and after that, the resemb lance to the quer y image and the image in the database will be obtained. Here are se v er al steps in f eature e xtr action phase: 1. Image Decomposition using Haar W a v elet In the first step , all Red, Green, and Blue component in database and quer y images are decomposed using 2 nd le v el Haar W a v elet. The results of this step are: appro ximate coefficient and v er tical, hor iz ontal and diagonal detail coefficients . After that, the appro ximate coefficient, hor iz ontal, and v er tical coefficient of Red, Green, and Blue components are combined. The combined appro ximate coefficient assign with 0.01, hor iz ontal with 0.008, and v er tical with 0.008 (e xper imentally obser v ed v alues). 2. Con v er t (LL, LH, and HL) of RGB to HSV The frequency sub bands which get from image decomposition steps (appro ximate (LL), hor iz ontal (LH), and v er tical coefficients (HL) where L denotes lo w frequency and H denotes high frequency) are con v er ted into HSV plane in order to e xtr act the color f eature . 3. Quantiz e HSV to (8,8,8) F or reducing the n umber o f colors , the color is quant iz e d using HSV color histog r am b y assigning 8 le v el each to Hue , Satur ation, and V alue components . So , the quantization will giv e HSV with 512 histog r am bins (8 x 8 x 8). 4. Compute the histog r am The last step is computing the nor maliz ed histog r am b y dividing with the total n umber of pix els . W a v elet-Based Color Histog r am on Content-Based Image Retr ie v al (Ale xander) Evaluation Warning : The document was created with Spire.PDF for Python.
1258 ISSN: 1693-6930 Figure 1. Flo w Diag r am of the Proposed Method After f eature e xtr action phase has been completed, the ne xt step is similar ity matching. The steps of similar ity matching consist of: 1. Similar ity computation with Distance Function After e xtr acting the f eatures of quer y image , the ne xt step to be tak en is computing the similar ity f eature of quer y image and all images in the database . The calculation is perf or med b y using histog r am intersection distance using the equation 1. Where |Q| represents the magnitude of the histog r am f or quer y image and |D| represents the magnitude of the histog r am representativ e image in database . d I D = P n i =1 min [ Q [ i ] ; D [ i ]] min [ j Q [ i ] j ; D [ i ]] (1) 2. Retr ie v ed Images The 10 most rele v ant images (with most similar histog r am) are sho wn as the result of retr ie v al. 3. Result and Anal ysis The e xper iment sho ws that WBCH using 2 nd le v el w a v elet giv es more precision than the others , including WBCH using the 1 st and 3 rd le v el w a v elet. The 2 nd le v el WBCH impro v es the a v er age precision of CBIR system f or 0.010. The compar ison of precision result betw een 2 nd le v el WBCH and the other methods is sho wn on T ab le 1 (W a v elet Based Color Histog r am / WBCH; Color Histog r am /CH; Color-T e xture and Color-Histog r am based Image Retr ie v al System / CTCHIRS; Color and T e xture F eatures f or Content Based Image Retr ie v al / CTIRS; The combination of color , te xture and shape f eatures using image and its complement / CTSIRS; Content based Image Retr ie v al System based on Dominant Color and T e xture F eatures / CTDCIRS). T ab le 1 sho ws the precision v alue of each categor y of the image and also the a v er age precision, while the sample of retr ie v ed images of e v er y categor y is sho wn on T ab le 2. The compar ison of precision and recall betw een 2 nd and 3 rd le v el WBCH are sho wn in figure 2, 3, 4, 5, 6, 7, 8, 9, 10, and 11. TELK OMNIKA V ol. 16, No . 3, J une 2018 : 1256 1263 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 1259 T ab le 1. Precision Result using Diff erent Methods Classes Categor y 3 r d Le vel WBCH 2 nd Le vel WBCH 1 st Le vel WBCH [12] CH [12] CTC- HIRS [9] CTI- RS [10] CTS- IRS [11] CTD- CIRS [14] 1 Afr ican P eople 0.836 0.856 0.650 0.720 0.680 0.750 0.540 0.562 2 Beach 0.441 0.468 0.620 0.530 0.540 0.600 0.380 0.536 3 Buildings 0.642 0.729 0.710 0.610 0.560 0.430 0.300 0.610 4 Buses 0.859 0.851 0.920 0.930 0.890 0.690 0.640 0.893 5 Dinosaurs 0.996 0.997 0.970 0.950 0.990 1.000 0.960 0.984 6 Elephants 0.678 0.723 0.860 0.840 0.660 0.720 0.620 0.578 7 Flo w ers 0.922 0.911 0.760 0.660 0.890 0.930 0.680 0.899 8 Horses 0.776 0.799 0.870 0.890 0.800 0.910 0.750 0.780 9 Mountains 0.958 0.946 0.490 0.470 0.520 0.360 0.450 0.512 10 F ood 0.462 0.485 0.770 0.820 0.730 0.650 0.530 0.694 A v er age Precision 0.757 0.777 0.762 0.742 0.726 0.704 0.585 0.705 T ab le 2. Sample Image Retr ie v al Results using 2 nd le v el WBCH Categor y Quer y Retrie ved Ima g es Afr ican P eople Beach Buildings W a v elet-Based Color Histog r am on Content-Based Image Retr ie v al (Ale xander) Evaluation Warning : The document was created with Spire.PDF for Python.
1260 ISSN: 1693-6930 Categor y Quer y Retrie ved Ima g es Buses Dinosaurs Elephants Flo w ers Horses TELK OMNIKA V ol. 16, No . 3, J une 2018 : 1256 1263 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 1261 Categor y Quer y Retrie ved Ima g es Mountains F ood Based on the precision and recall in figure 2-11, the w a v elet le v el 2 giv es slightly better precision than w a v elet le v el 3. As can be seen on se v er al categor ies , such as Afr ican P eople (Figure 2), Beach (Figure 3), Buses (Figure 5), Dinosaurs (Figure 6), Flo w ers (Figure 8), Mountains (Figure 10) and F oods (Figure 11), there are no significant diff erence betw een precision using w a v elet le v el 2 and w a v elet le v el 3. While the significant changes are highly visib le on the categor ies of Buildings (Figure 4), Elephants (Figure 7) and Horses (Figure 9). This e xper iment pro v es that w a v elet le v el 2 is mostly super ior than w a v elet le v el 3. 4. Conc lusion Based on the e xper iment conducted, it can be concluded that 2 nd le v el W a v elet Based Color Histog r am (2 nd le v el WBCH) is a better CBIR method compared to 1 st le v el WBCH, w a v elet te xture , color histog r am, and color co-occurrence matr ix. The a v er age precision of 2 nd le v el WBCH is 0.777, which impro v es the a v er age precision of 1 st le v el WBCH f or 0.010. The 2 nd le v el WBCH is also sur passing the a v er age precision of the 3 rd le v el WBCH. Since 2 nd le v el Figure 2. Precision and Recall f or Afr ican P eople Figure 3. Precision and Recall f or Beach W a v elet-Based Color Histog r am on Content-Based Image Retr ie v al (Ale xander) Evaluation Warning : The document was created with Spire.PDF for Python.
1262 ISSN: 1693-6930 Figure 4. Precision and Recall f or Buildings Figure 5. Precision and Recall f or Buses Figure 6. Precision and Recall f or Dinosaurs Figure 7. Precision and Recall f or Elephants Figure 8. Precision and Recall f or Flo w ers Figure 9. Precision and Recall f or Horses Figure 10. Precision and Recall f or Mountains Figure 11. Precision and Recall f or F ood TELK OMNIKA V ol. 16, No . 3, J une 2018 : 1256 1263 Evaluation Warning : The document was created with Spire.PDF for Python.
TELK OMNIKA ISSN: 1693-6930 1263 WBCH obtain promising result, w e kno w that this method can be applied in a CBIR system f or man y domains . F or future w or k, in order to impro v e the precision of image retr ie v al, shape can be included as the f eature to be e v aluated. Ref erences [1] N. P ar vin and P . Ka vitha, “Content based image retr ie v al using f eature e xtr action in jpeg domain and genetic algor ithm, Indonesian Jour nal of Electr ical Engineer ing and Computer Science , v ol. 7, no . 1, pp . 226–233, 2017. [2] W . Zukuan, K. Hongy eon, K. Y oungkyun, and K. J aehong, “An efficient content based image retr ie v al scheme , Indonesian Jour nal of Electr ical Engineer ing and Computer Science , v ol. 11, no . 11, pp . 6986–6991, 2013. [3] N. G. Rao , V . V . K umar , and V . V . Kr ishna, “T e xture based image inde xing and retr ie v al, IJCSNS Inter national Jour nal of Computer Science and Netw or k Secur ity , v ol. 9, no . 5, pp . 206–210, 2009. [4] M. J . Sw ain and D . H. Ballard, “Color inde xing, Inter national jour nal of computer vision , v ol. 7, no . 1, pp . 11–32, 1991. [5] X.-H. Han an d Y .-W . Chen, “Imageclef 2010 modality classification in medical image retr ie v al: Multiple f eature fusion with nor maliz ed k er nel function. in CLEF (Notebook P apers/LABs/W or kshops) , 2010. [6] H. T am ur a, S . Mor i, and T . Y ama w aki, “T e xtur al f eatures corresponding to visual perception, IEEE T r ansactions on Systems , man, and cyber netics , v ol. 8, no . 6, pp . 460–473, 1978. [7] C . Y ouness , O . Mohammed, A. Br ahim et al. , “Ne w method of conten t based image retr ie v al based on 2-d espr it method and the gabor filters , Indonesian Jour nal of Electr ical Engineer ing and Computer Science , v ol. 15, no . 2, pp . 313–320, 2015. [8] S . Y . Ir ianto , “Segmentation f or image inde xing and retr ie v al on discrete cosines domain, TELK OMNIKA (T elecomm unication Computing Electronics and Control) , v ol. 11, no . 1, pp . 119–126, 2013. [9] C .-H. Lin, R.-T . Chen, and Y .-K. Chan, “A smar t content-based image retr ie v al system based on color a nd te xture f eature , Image and Vision Computing , v ol. 27, no . 6, pp . 658–665, 2009. [10] G. Raghupathi, R. Anand, and M. De w al, “Color and te xture f eatures f or content based image retr ie v al, in Second Inter national conf erence on m ultimedia and content based image retr ie v al , v ol. 3, 2010, pp . 39–57. [11] P . Hiremath and J . Pujar i, “Content based image retr ie v al using color , te xture and shape f eatures , in Adv anced Computing and Comm unications , 2007. ADCOM 2007. Inter national Conf erence on . IEEE, 2007, pp . 780–784. [12] M. Singha and K. Hemachandr an, “Content based image retr ie v al using color and te xture , Signal & Image Processing , v ol. 3, no . 1, p . 39, 2012. [13] P . Suhasini, K. Kr ishna, and I. M. Kr ishna, “Cbir using color histog r am processing. Jour nal of Theoretical & Applied Inf or mation T echnology , v ol. 6, no . 1, 2009. [14] M. B . Rao , B . P . Rao , and A. Go v ardhan, “Ctdcirs: content b ased image retr ie v al system based on dominant color and te xture f eatures , Inter national Jour nal of Computer Applications , v ol. 18, no . 6, pp . 40–46, 2011. W a v elet-Based Color Histog r am on Content-Based Image Retr ie v al (Ale xander) Evaluation Warning : The document was created with Spire.PDF for Python.