A Novel 2D Texture Classifier For Gray Level Images

Transcription

1 2012, TextRoad Publication ISSN Journal of Basic and Applied Scientific Research A Novel 2D Texture Classifier For Gray Level Iages B.S. Mousavi 1 Young Researchers Club, Zahedan Branch, Islaic Azad University, Zahedan, Iran. ABSTRACT In this paper a new fuzzy ethod for 2D texture classification is proposed. A fuzzy rule base syste is designed based on coputed Geoetric Moents (GMs), which are rotation, scale and translation invariant, for each texture class. Although GMs are applied to classify texture, utilizing fuzzy inference syste (FIS) to ake decision, develops uch ore robust classifier. Where GMs are inputs and likelihood of belonging to one class is the output of fuzzy syste, the optiized threshold value is obtained by Genetic Algorith, to ake final decision. 90.8% classification rate proves the syste efficiency. KEY WORDS: Texture,Fuzzy Syste,Gray Level Iages,Classifier. 1. INTRODUCTION Texture analysis plays an iportant role in coputer vision and pattern recognition, and is widely applied in areas of industrial application; bio edical iage processing; reote sensing; and iage retrieval for classification, detection or segentation of iages based on spatial variation [1].Texture classification is a branch of texture analysis which results particularly well suited for the autoatic grading of products such as ceraic tiles, arble and granite tiles, parquet slabs, etc. Based on this fact, an increasing attention fro industry has recently eerged. Since in practical applications it is uncoon that texture iages are captured under invariant viewing conditions, it is of great iportance that texture classification be rotation, translation and scale invariant. Another issue in texture classification is about the role of colour [2]. Even if any approaches to texture analysis have been proposed in the last three decades, in ost cases such ethods are applied to grayscale iages. A coprehensive review of these techniques can be found in Petrou and García Sevilla work [3]. The ai of this work is to present a rotationally invariant descriptor for gray level textures. After studying various texture classification ethods, Geoetric Moent (GM), that is scale, position and orientation invariant, is chosen and fuzzy rule base syste is applied for classification. Geoetric invariant oent was first introduced in 1962 by Hu. It was derived fro the theory of algebraic invariant [6]. GM technique attepts to extract Rotation/ Scale/ Translation (RST)-invariant visual features. GM has been successfully applied in aircraft identification, texture classification and radar iages to optical iages atching [7]. Two-diensional oents of a digital iage f(x,y) is given as: p q x y f ( x, y) x y (1) The corresponding central oent is defined as: where: x x 00 y 10 & p q ( x x) ( y y) f ( x, y) y (2) (3) for p,q = 0,1,2,. The noralized central oent of order (p+q) is defined as: (4) *Corresponding Author: B.S. Mousavi, Young Researchers club, Zahedan branch, Islaic Azad University, Zahedan, Iran. Eail: bboosavi@gail.co 620

2 Mousavi, 2012 where: p q 2 1 (5) In particular, Hu defines seven values, coputed by noralizing central oents through the third order, that are scale, position, and orientation invariant. In ters of the central oents, these seven oents are [6]: (1) φ = η + η (6) (2) φ = (η η ) + 4η (3) φ = (η 3η ) + (3η η ) (4) φ = (η + η ) + (η + η ) (5) φ = (η 3η )(η + η )[(η + η ) 3(η + η ) ] + (3η η )(η + η )[3(η η ) (η + η ) ] (6) φ = (η η )[(η + η ) (η + η ) ] + 4η (η + η )(η + η )] (7) φ = (3η η )(η + η )[(η + η ) 3(η + η ) ] + (3η η )(η + η )[3(η + η ) (η + η ) Fuzzy sets theory provides a fraework to aterialize a fuzzy rule-based syste which contains the selection of fuzzy rules, ebership functions, and the reasoning echanis. Such systes have been applied to any disciplines such as control systes, decision aking and pattern recognition [4], and in this paper it is supposed that such a syste could overcoe the coplexity of the texture classification probles, ainly are known as: variable conditions. Fuzzy logic if-then rules are fored by applying fuzzy operations to these ebership functions for given inputs. The resulting output ebership functions are added together using desired weights yielding a sort of probability function. This function can then be used to estiate the expected value of the output variable. Madani type is one the ost coonly used fuzzy inference ethod which is eployed in this study as well [5]. In this paper, accurately designed Madani fuzzy rule base syste is proposed for gray scale texture classification. GMs are eployed as inputs and different texture clusters are outputs. More details about various steps of syste designing and its characteristics are presented in following sections. The reinder of paper is organized as follows. Proposed ethod is analyzed in section 2. Section 3 and 4 present obtained result and conclusion, respectively. 2. PROPOSED METHOD After various investigation GMs are found as one of the ost reliable ethods to texture classification. Table 1 shows obtained GM values for six different texture depicted in Fig. 1. These iages are fro Outex [8] texture iage database. The nuber of each category and the nuber of texture in that category (in parentheses) could be seen at the botto of each iage. It should be noted, GM values in table 1 are the absolute values of log of results, obtained by applying equation (6). Using log, dynaic range would be reduced and the absolute values avoids having to deal with the coplex nubers that result when the log of negative oent invariant is coputed. As it is ore difference between 5 to 7 values for different categories, these values are chosen to classify. Instead of using crisp threshold, 621

3 fuzzy rule base syste is applied. This fuzzy syste is a Madani type, with three inputs ( 5-7 ) and one output which is the likelihood of belonging to special category for any input texture. Fig.1. Texture iages fro Outex database Table 1. GM values for textures shown in Fig. 1. GM Iage 001(1) Iage 002(12) Iage 003(6) As an exaple, to show the robustness of proposed ethod, a syste is designed to select and classify the first category in Outex database. Subtractive clustering [9] is applied on input space (contain 600 various texture iage) to decide on the nuber of ebership functions (MF's) and rules. Utilizing the obtained three clusters inforation and experiental knowledge, input and output MF's are designed. The seantic eaning is assigned to each cluster for better understanding. The achieved rule in texture classification FIS is: IF input is Z, THEN output is Z where Z { Belonging, Rather-Belonging, Not Belonging } MF's of inputs and output is depicted in Fig. 2. To achieve the crisp output, centroid ethod is chosen, which is the ost widely used one aong all defuzzification approaches [5]. The output is the texture-likelihood, between 0 to 1. This value reveals the probability of belonging to desired texture group, for each arbitrary input texture iage saple. (a) (b) 622

4 Mousavi, 2012 (c) (d) Fig.2. (a-c) input and (d) output MF's. To ake final decision a suitable threshold value should be selected. To find an optiized value, genetic algorith (GA) is applied. GA is the ost extended group of evolutionary technique known, which rely on the use of a selection, crossover and utation operators [10]. The threshold is the chroosoe of the GA, whose fitness function attepts to axiize the output likelihood for texture belonging to desired texture group. Over 100 randoly selected texture saples, the obtained threshold is It eans that the texture with likelihood ore than 0.78 are regarded as texture belonging to one class. 3. RESUNLTS AND DISCUSSI To show the syste efficiency and evaluate its perforance, the proposed ethod is applied on Outex texture database. This iage database contains a large collection of textures, both in for of surface textures and natural scenes. The collection of surface textures exhibits well defined variations to a given reference in ters of illuination, rotation and spatial resolution [8]. A large collection of texture classification, retrieval and segentation probles, both supervised and unsupervised, is constructed using the iage database. The diversity of the surface textures provides a rich foundation for building the probles. For exaple, in addition to standard texture classification, probles of illuination/rotation/resolution invariant texture classification, or their cobinations, are also available. Different isclassification cost functions and prior probabilities of classes are also incorporated. The described syste in previous section is applied on 1000 texture iage of Outex database, suite ID fro Outex_TC_0000 to Outex_TC_0009, with window size fro 32x32 to 128x128. The designed fuzzy inference syste showed successfully 90.8% correct classification rate over these iage textures. Considering wide range of various iage textures, specially textures with different size and rotation angles, it can be said that proposed syste is nearly rotation and size invariant. This feature is valuable characteristic for a texture classifier. 4. Conclusion and Future Works In this paper we presented a novel fuzzy inference syste to classify texture. GMs were used as syste inputs to design a rotation, scale and translation classifier. As the syste output shows likelihood of belonging to the special class, an optiized threshold value was achieved using Genetic Algorith. Utilizing fuzzy ethod and selecting optiizing threshold value, a reliable classifier was designed, whose correct classification rate (90.8%) proved this idea. Adding colour inforation to this syste, for colour texture classification, is our next ai. 623

5 5. REFERENCES [1] F. Lahajnar, S. Kovacic, "Rotation-invariant texture classification", Pattern Recognition Letters 24 (2003) [2] F. Bianconi, A. Fernández, E. González, D. Caride, A. Calviño, "Rotation-invariant colour texture classification through ultilayer CCR", Pattern Recognition Letters 30 (2009) [3] Petrou, M., García Sevilla, P.G., Iage Processing. Dealing with Texture. WileyInterscience. [4] K. Tsuda, M. Minoh, K. Ikeda, "Extracting straight lines by sequential fuzzy clustering", Pattern Recognition Lett. Vol. 17, pp , [5] S.N.Sivanandu, S.Suathi, S.N.Deepa, Introduction to Fuzzy logic using MATLAB, springer-verlag Berlin Heidelberg, [6] M. K. Hu, Visual pattern recognition by oent invariants, IRE Trans. Inforation Theory, vol. 8, 1962, pp [7] J.Flusser, Moent Invariants in Iage Analysis, Proc. Of World Acadey of Science, Engineering and Technology, Vol. 11, [8] [9] Agus Priyono, Muhaad Ridwan, Generation of Fuzzy Rules With Subtractive Clustering, JurnalTeknologi, Vol. 43, No. D, pp , [10] S.N.Sivananda S.N.Deepa, Introduction to Genetic Algoriths, Springer-Verlag Berlin Heidelberg, pp ,