Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal    of Engineering Research ...
Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal    of Engineering Research ...
Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal    of Engineering Research ...
Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal    of Engineering Research ...
Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal    of Engineering Research ...
Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal    of Engineering Research ...
Upcoming SlideShare
Loading in...5
×

Cu25572577

106

Published on

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

0 Comments
1 Like
Statistics
Notes
  • Be the first to comment

No Downloads
Views
Total Views
106
On Slideshare
0
From Embeds
0
Number of Embeds
0
Actions
Shares
0
Downloads
3
Comments
0
Likes
1
Embeds 0
No embeds

No notes for slide

Cu25572577

  1. 1. Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.572-577 An Art Images Retrieval System Based On Similarity Descriptor Concepción Pérez de Celis1, Thania Felix1, Maria Somodevilla1 and Ivo Pineda1 1 Benemérita Universidad Autónoma de Puebla, Puebla, México,ABSTRACT Last decade Content Based Image semantic interpretation is beyond the scope of thisRetrieval (CBIR) has captured attention due to work.the increase of knowledge of multimedia The use of light in a painting is closely related todatabases in the different areas. In this paper we the color and composition, the degradation of lightpresent an image retrieval art system, which was helps to create an atmosphere and depth effects. Thedeveloped to query and to explore a collection of use of light is an important feature by itself whenart images by QBE, considering similarity criteria classifying styles of painting in question, for instancelike: color palette, use of light and the technique lighting in Impressionism is considered as aused. These similarity criteria are analyzed by protagonist. When speaking of paint composition andextracting low-level features such as color, texture spatial organization is referred to the line but also theand spatial relationships. stroke, i.e. if the shapes that constitute the images created by the painter are produced by drawing orKeywords: CBIR, color histograms, Art created by glazing such as Renaissance paint usingdatabases the technique of sfumato. Texture defines the quality of the surface in terms of its appearance; the actual texture refers toI. INTRODUCTION the physical characteristics of the stroke. The impasto Currently a number of museums and is the application of pigment so little fluid creating agalleries have been incorporated into the web as a dense layer. A painting has also a virtual textureservice to the public display of their collections. One achieved by the impact made by the painter who canof the advantages of having such tools would be the give the rough appearance being different from itspossibility that visitors to a museum and students of actual texture. Finally, color is perhaps the bestart history could have access to visit the collection known formal element, when speaking of color it isconsidering their interests. Suppose, for example that considered three elements: hue, saturation anda person after visiting a museum is interested in brightness or intensity. In painting it is common tofinding paintings similar to those seen during the speak of the color palette specially when it is relatedcourse of a given exposure, this implies the existence with an art movement because colors represent aof mechanisms that allow dynamic browsing consider distinctive pattern inside paintings, for example thethe similarity criteria selected by the user i.e. same art color palette of Picasso and Braque Cubist paintersmovement, same artist, time, among others. have muted colors (ocher , dark green and gray Therefore there are two fundamental variety).problems when making content retrieval of imagesart: the extraction of representative features from thecontent and definition of a similarity measure. In thispaper IMTHAC system is presented, which uses low-level features to extract content in images todetermine the similarity between two pieces ofpictorial art through color palette used, light insidethe paints, lines, texture, composition and depth.I.1 PICTORIAL FEATURES FROM THE ART HISTORYPERSPECTIVE Figure 1. Targeted search example, in this case theThe formal approach assumes painting styles [1] [2] searched images are of same object with differentwhere the observer understands the art in formal color characteristics.terms of color, shape (linear or pictorial) and lighting,in addition to the iconographic content. Each of the I.2 ART IMAGES RETRIEVAL SYSTEMS (AIR)formal elements for clarity in the explanation of the Retrieval systems for art images usuallyproposed approach corresponds to a category. Thus, have at least three components: formal analysis,four categories will be under analysis: Light, Line, comparison of the paintings formal aspects and styleTexture and Color. Categories like iconographic classification. There are several documented searchanalysis are not considered, since this requires models for art images, some of them are: [13] [14]identification and connotation of the objects using a 572 | P a g e
  2. 2. Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.572-577I.2.1 SEARCH BY ASSOCIATION global distribution of color and (2) for local For this model all the images in the database distribution or color regions. The relevant differenceshould be analyzed, whereas similarity criteria for a between the two techniques is by indexing globalgiven image or sketch should be considered. This distribution can only make comparisons with overalloften involves iterative refinement of the search. image while local distribution enables the search for localized regions within the image. Both techniquesI.2.2 TARGETED SEARCH have their advantages and they are designed to solve The goal, in this type of search, is to find a different types of queries. The overall indexing byspecific image. The search can be for an exact copy color is more useful when the query comprises anof the image given as an example or target; it can also image as an example.search for another image of the same object. Suppose Since 1980 several algorithms which usea user has a specific image and wants to find images color extraction from images have been developedwith the same objects such as another painting [8]. The most common methodology for comparingMarylin series of Andy Warhol (1962), an example the contents of a query image with the rest of theof this type of search is shown in Figure 1. For images in a database is to compare their colortargeted searches it should specify the target image’s histograms In [9], Jain and Vailaya considered colortype by saying return exact image or object with space RGB and obtained a histogram for eachchromaticity, texture, etc.. channel of the image, the similarity was measured using the Euclidean distance to compare each bin ofI.2.3SEARCH BY CATEGORY the histograms. Jeong [10] instead of using three The goal is to find an arbitrary histograms for each of the RGB channels trying torepresentative corresponding to the specific class of create bins that contained the information of the threethe sample image. In the category search, the user channels and therefore the number of bins wasmay have available a set of images including greater. Some tests were also carried out with theadditional images of the same class. For example, HSV model. Other two examples of systems usingsearch for other paintings by the same artist or same the global color distribution but focused on Art arestyle of painting. QBIC [11] and PICASSO [12]. Working with art images, we can also III. SYSTEM PROPOSALidentify two types’ intervals in the range of The following sections provide a briefmeanings: sensory gap and the range of content or overview of the work performed to support themeaning (semantic gap). Sensory Interval relates to selection of the methodology used to develop thediversity between the object in the real world and that search model implemented by association,seen in the paint under analysis [3]. For digital image considering elements of similarity such as the coloranalysis has to consider the use of color, the palette, spatial distribution of color, lighting, andexistence of edges, brightness and texture for it texture.considering the statistical properties of images. Forthat matter, the range of content or meaning refers to III.1 EXTRACTING THE COLOR PALETTE FROMdifference between the immediate interpretation of IMAGESpictorial information in all its different forms and the As the main similarity criterion of similarityinterpretation that follows from a formal description the color palette used is considered. We understandof the object [4]. the whole color palette of dominant colors in a painting that may be associated with age, author or In this paper, the sensory range to define the style of painting. Histograms are used to extract thesimilarity criteria between images and the system palette of color. The histogram of an image isdeveloped using the search pattern by association is defined by the probability mass function of the imageconsidered. intensities. This extends for color images to capture the joint probabilities of the intensities of the three In subsequent sections we will first briefly color channels. Formally, a color histogram isreview previous work related to the search patterns defined in (1),by association particularly those who consider the H A,B,C (a, b, c) = N * Prob (A=a, B=b, C=c) (1)color as a criterion of similarity. After this, we where A, B and C represent color channelsdescribe our proposal considering as similarity (R, G, B ó H, S, V) and N, is the number of pixels inelements the color palette, spatial distribution of the image. Computationally, the color histogram iscolor, texture and lighting. Finally, the results obtained by discretizing the colors of an image andobtained to date and ongoing work is presented. the number of pixels of each color [15]. To generate the color histograms, a color space whichII. RELATED WORK describes mathematically how the colors can be There are basically two techniques for represented, is needed. From existing color spacesretrieval (indexing) images by color: (1) for the just two of them were utilized in this work: the RGB 573 | P a g e
  3. 3. Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.572-577color space, the most widely used by the vast was added allowing the precision increasesmajority of video and photographic cameras, and significantly.HSV resembles the way in which the human senses Some improved techniques for colorcolors and how it is capable of separating color histograms including spatial characteristics arevalues. Given the characteristics of the images to be reported in the literature. One of these proposes theanalyzed, HSV color space was chosen for this work. construction of the histogram by creating concentricAdditionally, since in the images analyzed, colors are circles in the image [6] to which is added a factor,generated by the juxtaposition of tones, quantization considering the histogram of the first circle, which isprocess is used to standardize the range of colors to the one corresponding to the center of the image asbe treated. The quantization process aims to reduce that of greater weight. Lei[5] proposes to determinethe number of different colors used in a visual image for each color ci, from the quantized color set C, thewithout losing information. Another advantage to position of its center of mass normalized. Theyreduce the number of colors is to improve the consider also, two weight coefficients α and β for theprocessing speed. user’s level of interest by color or spatial The quantization method proposed by Zhang information. In [7] Cinque et al. present SpatialLei, Fuzong Lin and Zhang Bo [4] was modified, Chromatic Histograms (SCH), which combinebased on a quantization to 36 colors using non- information on the position (location) of the pixels ofuniform HSV color space. This quantization works similar color and their arrangement within the imagewell with everyday images, however the results of with the classic color histogram. For each color in theour images were not satisfactory so instead of making quantified image, the percentage of pixels having thea non-uniform quantization, it was necessary to make same color is calculated, and spatial informationa uniform quantization to 56 colors, as the color summarized in the relative coordinate of thepalette, which is handled in art. The change made barycentre of the spatial distribution besides itswas to consider only 12 colors instead of 7 colors, corresponding standard deviation. Severaltaking into account the color wheel used by artists. experiments were carried out to test the algorithmsOnce the images were quantized to 56 bins, their behavior described in the preceding paragraphs whenhistograms were obtained. Comparisons measure dealing with art images. These experiments allowedbetween histograms is shown in (2), to decided whether include or not the barycentre ofi Ri Si (2) the different colors found in the image and its standard deviation. where Ri and Si are the color histograms The proposed algorithm is as follows:and, i is the number of pixels in the image. Figure 2 For an image I already quantized to 56 bins,shows the dynamics of the process done. Pk (I) := { (x,y) I : [x,y]=k} (3) as the set of pixels in I whose color is k. It is also calculated hi(k) corresponding to the percentage of pixels of color K that are in I, PkI hI  k   (4) nm with nxm the total number of pixels of the image I, and compute for each color k its barycentre,, bI  k    xk , yk  where:  1 1Figure 2. The query image in RGB space is xk  x (5)converted to HSV space, then the HSV image is n Pk  x , y Pkquantized to 56bins, its color histogram is extracted  1 1and compared with the histogram of the existing yk  y (6) m Pk  x , y Pkimages in the database, returning the similar imagesin color, and allowing the user to determine the The barycentre gives an idea of the pixels’percentage of similarity. position having the same color; however it is still a rough approximation of the spatial properties of theIII.2 SPACE-COLOR ALGORITHM image pixels. Since several pixels’ arrangements of Spatial information was included in order to the same color can have closer barycentres, in orderavoid losing image color distribution information in to improve information about the spatial propertiesthe image, and to avoid false positives. In particular, and regions with higher concentration of pixels ofbesides the color its position in the query image plane one color, then the standard deviation of the pixels 574 | P a g e
  4. 4. Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.572-577for Pk (I) is obtained. Let p be a pixel in any Pk (I) If | RIq|≤ S, the efficiency is reduced as far aswith relative coordinates (xp, yp) the standard traditional recall whereas that |REq|≥ S efficiencydeviation is then defined in (7), corresponds to the precision.  d  p, b  k  1 2 l  I Pk pPk (7) It requires a new function of similarity (8)between two images, as the function used in [7], ifthe barycentre and the standard deviation in thecalculation of color histogram is included, with cf s  Q, I    min  hQ  i  , hI  i    c i 1  2  d  bQ  i  , bI  i   min  Q  i  , I  i         2 max  Q  i  , I  i     (8) the number of bins resulting from thequantization, Q being the query image and R anyother image that is in the database.III.3 OTHER CHARACTERISTICS UNDER ANALYSISOther features that can improve the results ofsimilarity between images were also considered aslightness component of color images, dark pixels andtexture. For texture, the method of gray level Figure 3. Diagram of color and texture distribution.difference [16, 17] (GLDM)) was implemented. Thetexture is used in combination with the color Figures 4, 5, 6 and 7, show system responsesdistribution. In Fig. 3, an implemented algorithm to search for similar images to the query image todiagram for searching images by similarity in color each of the algorithms implemented for art imageand texture is shown. retrieval by association. For all cases the top ten most relevant images are shown besides the similarityIV. EXPERIMENTAL RESULTS percentage of each of the resulting images with the A database with 1300 art images from image query. In this case, a limit was established asdifferent painting styles: Renaissance, Classicist, much as 50% similarity reaching an efficiency ofMannerist, Baroque, Impressionism, Post- 81%. Additionally, the Fig. 8 presents the MAPImpressionism and Cubism was use to test the (mean average precision) results from different testsalgorithms proposed. These images were classified performed on the images database grouped by theby an expert. The results were evaluated by pictorial style of the desired image.efficiency measure (Efficiency of Retrieval or FillRatio) defined by Methre et al. and they also wereused for similarity in the color image retrieval byBarolo et al. [7]. This measure proposed to be S thenumber of relevant parts that the user wants toretrieve from an image given as example, IQR thetotal number of relevant images, and REq the numberof relevant images retrieved in the short list. Themeasure of effectiveness is given in (9),  RI  RE  RI  S q q if q  Rq I S   I  Rq  Rq E  E if RI  S q  Rq  (9) Figure 4. Example of query using color algorithm 575 | P a g e
  5. 5. Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.572-577 V. CONCLUSIONS AND FUTURE WORK The research presented focuses primarily on the color palette of art images being enriched with texture and brightness detected in them. It also considers the spatial distribution of color thereby allowing not only retrieve images that are similar in color palette and lighting but can additionally retrieve images similar in content. The results so far are encouraging and we plan to incorporate searching by category to enrich the search model. In particular, searching by a specific artist or images corresponding to a particular pictorial style is under consideration.Figure 5. Example of query using color distributionalgorithm REFERENCES [1] E. H. Gombrich. Arte Percepción y Realidad (Paidos Ibérica, 2003). [2] E. Panofsky, La perspectiva como forma simbólica (Tsquets, 1999). [3] G. T. Smeulders Content-based Image Retrieval: An Overview, Survey on content- based image retrieval. In: Emerging Topics in Computer Vision (Prentice Hall, 2004). [4] A. W. Smeulders Arnold, L. Hardman Lynda, A. T. Schreiber, J. M. Geusebroek An integrated multimedia approach to cultural heritage e-documents, ACMFigure 6. Example of query using color and light Workshop on Multimedia Informationdistribution algorithm. Retrieval, ACM Press, 2002. [5] Z. Lei, L. Fuzong, Z. Bo. A CBIR method based on Color-Spatial feature, Proc. of the IEEE Region 10 Conference, Cheju , South Korea, 1999, 166 – 169. [6] W. Xiaoling, M. Hongyan. Enhancing Color histogram for Image Retrieval, Proc. of the International Workshop on Information Security an Applications, Qingdao, China, 2009, 622-625. [7] L. Cinque, G. Ciocca, Color-based image retrieval using spatial-chromatic histograms.Figure 7 Example of query using color and texture Image and Vision Computing, 19(2), 2001,distribution algorithm. This case only shows five páginas 979-986.results as they are the most similar to 50% similarity. [8] J. R. Smith, John and S. Chang, Tools and Techniques for Color Retrieval. Electronic Imaging:Science and Technology-Storage Retrieval for Image and Video Database IV. San Jose: IS&T/SPIE, 1996, 1-12. [9] J. A. K., and A. Vailaya. Image Retrieval Using Color and Shape. Great Britain: Elsevier Science LTD, 1995. [10] J. Sangoh. Histogram-Based Color Image Retrieval, Psych221/EE362 Project Report, Stanford University, 2001. [11] IBM QBIC at the Hermitage Museum. http://www.hermitage-museum.org [12] A. del Bimbo, M Mugnaini, P. Pala and F. Turco, VisualQuery by Color Perceptive Regions, Pattern Recognition Society, 31(9), 1998, 1241-1253.Figure 8. MAP measure results implemented in the [13] P. L. Stanchev, D. Green Jr., B. Dimitrov,seven classes. Content based image retrieval techniques – 576 | P a g e
  6. 6. Concepción Pérez de Celis, Thania Felix, Maria Somodevilla, Ivo Pineda / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.572-577 Issues, analysis and the state of the art, Some Issues in the Art Image Database Systems, Journal of Digital Information Management, 4(4) 2006, 227-232. [14] V. V. Kumar, , N. Gnaneswara Rao, and A.L. Narsimha Rao Reduced Texture spectrum with Lag value Based Image Retrieval for Medical Images, International Journal of Future Generation Communication and Networking, Vol. 2, No. 4, December, 2009, 39 48. [15] J. K. Kim and H. W. Par, Statistical Textural Features for Detection of Microcalcifications in Digitized Mammograms, IEEE Transactions on medical imaging. 1999. [16] Jong Kook Kim and Hyun Wook Par, Statistical Textural Features for Detection of Microcalcifications in Digitized Mammograms, IEEE Transactions on medical imaging. Vol 18(3),1999, 231-238 [17] A. C. Silva Algoritmos para Diagnóstico Assistido de Nódulos Pulmonares Solitários, Ph.D. Thesis, PUC-Rio, 2004 577 | P a g e

×