An Innovative Technique of Texture Classification and Comparison Based on Long Linear Patterns

The present paper proposes a method of texture classification based on long linear patterns. Linear patterns of long size are bright features defined by morphological properties: linearity, connectivity, width and by a specific Gaussian-like profile whose curvature varies smoothly along the crest line. The most significant information of a texture often appears in the occurrence of grain components. That’s why the present paper used sum of occurrence of grain components for feature extraction. The features are constructed from the different combination of long linear patterns with different orientations. These features offer a better discriminating strategy for texture classification. Further, the distance function captured from the sum of occurrence of grain components of textures is expected to enhance the class seperability power. The class seperability power of these features is investigated in the classification experiments with arbitrarily chosen texture images taken from the Brodatz album. The experimental results indicated good analysis, and how the classification of textures will be effected with different long linear patterns.


INTRODUCTION
Texture classification is an image processing technique by which different regions of an image are identified based on texture properties. This process plays an important role in many areas such as industrial automation, biomedical image processing, Content Based Image Retrieval and remote sensing application. In spite of the importance of textures in many areas of image processing, there is no universally accepted definition for the texture. We prefer to adopt the definition suggested in [1] , because of its generality and it is given as follows: "The notion of texture appears to depend upon three ingredients: (i) some local 'order' is repeated over a region which is large in comparison to the order's size, (ii) the order consists in the nonrandom arrangement of elementary parts, and (iii) the parts are roughly uniform entities having approximately the same dimensions everywhere within the textured region". This definition explains that the texture is characterized not only by gray value at a given pixel, but also by the gray value pattern in the surrounding pixels. The texture has both local and global meaning, in the sense that it is characterized by the invariance of certain local attributes that are distributed over a region of an image [2] . To design an effective algorithm for texture classification, it is essential to find a set of texture features with good discriminating powers. Most of the textural features are generally obtained from the application of a local operator, statistical analysis, or measurement in a transformed domain. Generally, the features are estimated from co-occurrence matrices, Law's texture energy measures, Fourier transform domain, Markov random field models, local linear transforms etc. A number of texture classification techniques are reported in literature [3,4,5,6] . The wavelet methods [3,6,7] offer computational advantages over other methods for texture classification and segmentation. In [8] , Haralick features [9] are obtained from wavelet decomposed image yielding improved classification rates. In [10] , texture features are characterized by considering intensity and contextual information obtained from binary images. The conditional co-occurrence histograms are computed J. Computer Sci., 3 (8): [633][634][635][636][637][638]2007 . 634 from the intensity and binary images. To obtain binary images the fixed thresholds were used. It is evident that the context or the position information of a pixel in an image is very important for the purpose of classification.
In this, paper, we propose a novel scheme of texture classification based on sum of occurrences of grain components in long linear patterns (LLP) with different orientations. This paper is organized as follows: In section 2, we briefly review theory of linear patterns. The methodology of feature extraction and distance function are evaluated in section 3. The analysis on texture classification based on experimental results is presented in section 4. Concluding remarks are given in section 5.

LONG LINEAR PATTERNS (LLP)
The linear elements considered in the present paper are long one-dimensional line elements or patterns or structuring elements. The one-dimensional LLP play a significant role in many image processing operations, such as segmentation, edge detection, classification etc. Moreover line structuring elements are more suited for many morphological operations. In some image processing applications, the square, hexagon, and octagon patterns are used. However the above patterns can be easily decomposed into two, three and four line segments respectively [11,12] . The advantage of using one dimensional line pattern segments instead of Ndimensional structuring elements is a reduction in the computational complexity.
In the present paper LLP on the texture images are computed basically by using Bresenham's line algorithm [13] . Bresenham [13] published an algorithm to draw a line segment of any size and of any orientation on a plotter, which could draw horizontal, vertical and diagonal lines. The algorithm combines small portions of these lines to form a line pattern of any orientation. In image processing, Bresenham lines are formed by steps in the eight cardinal directions of the grid. But true line patterns can be formed by using Bresenham's line algorithm, for the lines on X axis, Y axis and X=Y axis, i.e. the lines with 0 0 , 180 0 , 90 0 , 270 0 ,45 0 ,135 0 ,225 0 ,315 0 . For all other orientation, the shape of the LLP appears as steps of a stair case, known as Jaggies.
The LLPs of size 1x11 with 0 0 , 15 0 , 30 0 and 45 0 orientations are shown in the following Fig. 1(a), 1(b), 1(c) and 1(d) respectively. The following figures clearly indicate that the LLP length, defined by an integer number of pixels, depends on the degree of orientation. Each degree of orientation of the LLP will have a (a) different set of lengths. The present paper assumes the orientation of LLP as triangular shape shown in Fig.2, but not as a semi circle. This is one of the reasons why J. Computer Sci., 3 (8): [633][634][635][636][637][638]2007 . 635 length of the LLP is reduced for orientations other than X axis and Y axis. For example the LLP 1x11, is having length 11, 11, 9 and 9 for 0 0 , 15 0 , 30 0 and 45 0 orientations respectively. The length is decreased for 30 0 and 45 0 orientation because of the triangular orientation as shown in Fig.2. Where as the exact length of 11 pixels are resulted for the 0 0 and 15 0 orientation, as shown in the Fig.1 (a) and 1(b) respectively. This is because 0 0 orientation falls on X-axis and 15 0 orientation is very closer to X-axis.

METHODOLOGY
The present paper is not concentrated in studying the frequency of occurrence of one dimensional LLP on the texture image because their frequency occurrences will be minimum (most of the times one digit). The classification may become a problem if the frequency count is low. Moreover the most significant information of a texture appears in the occurrence of grain components. That's why the present paper had chosen the sum of occurrence of grain components of LLP as feature extraction. For this one dimensional LLP's of different sizes 1x11, 1x13, 1x15, and 1x17 are chosen. The above LLP's are rotated for every 15 0 from 0 0 to 90 0 . The change of topology on LLP with respect to summation of number of occurrences, of grain components are counted, as shown in Fig.1. This experiment is carried out on thirteen Brodatz textures and is listed from Table 3 to Table 6. In order to classify the textures a distance function is used in the present paper. The distance function is calculated in the following way.
Distance between Textures with same orientations.
where GX 1 o represents the total number of occurrences of grain components by X i o , of texture T k. And D (T 1 , T 2 ) is the absolute overall difference between the textures T 1 and T 2 .

RESULTS:
The above scheme of classification is applied on randomly chosen 13 Brodatz textures as given in Table 2. Here on wards the texture number is presented instead of texture name. The Tables 3, 4 The textures that differ with a distance threshold factor of'd' can be considered as one class. That is two or more textures can be placed into one class 'C' if each texture differs with all other textures in the group by a distance of less than or equal to d, as specified below. C = {T i , T i+1 , T i+2 , ----, T n }, this is true if and only if for all textures, D (T i, , T j ) <= d, where i, j are 1 to n and i j.
The following analysis on classification of textures for all LLP's has been done with the same distance threshold value 20. Careful analysis of Table 7, of LLP 1x11 reveals the following texture classification for a unique distance threshold value, 20.   T1  T2  T3  T4  T5  T6  T7  T8  T9  T10  T11  T12 T1  T2  T3  T4  T5  T6  T7  T8  T9  T10  T11  T12  T13   SE0   0   289, T1  T2  T3  T4  T5  T6  T7  T8  T9  T10  T11  T12  T13   SE0   0   330, T1  T2  T3  T4  T5  T6  T7  T8  T9  T10  T11  T12  T13   SE0 T1  T2  T3  T4  T5  T6  T7  T8  T9  T10  T11  T12  T13 where d i =20 The study of distance function  Table 9 of LLP 1x15 depicts the following texture classes for a distance threshold value of 20. The analysis about distance functions of LLP concludes that the number of overlapping texture classes will be more by increasing the linear element size. The texture groups or classes are very concise by LLP of size 1x11. The sum of the grain components of large linear patterns 1x11, 1x13, 1x15 and 1x17 as listed in the Tables 3,4,5 and 6 respectively indicates a decreasing trend from 0 0 to 45 0 and an increased trend from 45 0 to 90 0 of grain components. This clearly reflects the geometrical property on linearity that is reflection about the line X = Y. From this the present paper concludes that, it is not necessary to compute the orientations of linear patterns from 90 0 to 360 0 , as they can be counted merely by reflection of X axis and Y axis.