Published on

  • Be the first to comment

  • Be the first to like this

No Downloads
Total views
On SlideShare
From Embeds
Number of Embeds
Embeds 0
No embeds

No notes for slide


  1. 1. Ranjita Mishra / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 3, Issue 2, March -April 2013, pp.237-241 Scene Image Analysis using GLCM & Gabor Filter Ranjita Mishra#1 # Electronics & Telecommunications Department, Silicon Institute of Technology , BPUT, OdishaABSTRACT In this paper, images of real- world is the energy concentrated in the low spatialnatural scenes and manmade structures of frequency components contrary to the behaviourdifferent depth are taken. With increase in image shown by manmade structures. Therefore it maydepth , roughness increases in case of man-made be perceived that images of distant manmadestructures whereas natural scene images objects and close natural scenes have similarbecome smooth, thus reducing the local energy content. Similarly, the energy content ofroughness of the picture. Such kind of specific distant natural scenes and close manmade structuresarrangement produces a particular spatial are comparable.pattern of dominant orientations and scales We consider here a fundamental problem ofthat can be described using Gabor filter as it computer vision, i.e. enabling computers to see thegives the local estimate of frequency content in an way we see things. We in future wish our machinesimage. Here various techniques are used i.e. grey would match the capabilities of human vision. It‟slevel co-occurrence matrices (GLCM), Gabor interesting to note that, every second we receivefilters, combined GLCM and Gabor filters. Here tremendous amount of visual data and almostthe real scene images are classified in four classes unconsciously we process this information verysuch as near natural, near manmade, far natural quickly. Classifying an object as table, a ball, or aand far manmade .Gabor filter only classify into scene as mountain or river is pretty trivial for us. Welow energy and high energy scenes. So the can in fact process amazingly more complexcombination of Gabor filter and GLCM are used information. It‟s a well known fact that roboticfor classification in to four classes. In the vision compares miserably with our eyes. Here, weproposed method i.e. the combination of Gabor intend to make a start towards our goal byand GLCM, first Gabor and PNN is used for considering a very trivial problem by the standard ofclassification between two groups high energy human vision and that is scene classification. Given(such as near natural & far manmade) and low an image of the scene we wish to classify it as say aenergy( such as near manmade & far natural) mountain, forest, city, street etc. Generally, aand then the GLCM and PNN is used for learning based approach is used to solve problems ofclassification of subgroups. this nature. A training set is initially created which would contain representative images from allKeywords— Ga bor fi l t er s , G L CM, scene categories that we need to classify. Now theseclassification, images are manually labelled to the class they belong as perceived by the human. Now a learningI. INTRODUCTION algorithm is employed, which basically is a strategy Images of natural scenes and manmade to enable us to come up with parameters whichobjects exhibit a particular behaviour with would characterize an image for doing therespect to varying depth [9]. Close-up views of classification task. Now if a random image is givenmanmade objects generally contain large flat as an input, on basis of parameters already identifiedsurfaces and homogeneous surface areas are the machine would try to classify the image. This inlikely to be replaced by objects like ,walls ,doors, essence a generic way in which learning algorithmswindows etc thereby increasing the global work, i.e., by learning from a huge set of data androughness of the picture. On the other hand , then using this learned information to makenatural scene images show a radically different predictions about successive inputs.behaviour with respect to mean-depth when In this work we are trying to recognizecompare with manmade objects. Close-up views the scenes of four different categories namelyof natural scenes are usually textured with near-natural, near-manmade, far-natural, & far-regular repetition of some specific pattern. But as manmade. In our work different methods are used forthe depth of such image increases small grains of features extraction like Gabor filter, GLCM, andclose-up view are indiscernible, because there is combination of Gabor filter and GLCM. Gabor filtera change in spatial frequency of the image. As a can classify the scene images into two categories basedresult natural structures become larger and on Gabor Energy. i.e. Low energy and High energysmoother giving rise to homogeneous regions i.e. scene images. Near-manmade and far-natural scenelow in roughness. Thus in natural scene, more images have low energy whereas, Near-natural & far-we increase the mean depth of the scene, more manmade scene images have high energy. So Gabor 237 | P a g e
  2. 2. Ranjita Mishra / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 3, Issue 2, March -April 2013, pp.237-241Filter can classify these scene images into two groups receptive field. Its ellipticity and herewith thebut it cannot classify in two four different class. ellipticity of the receptive field ellipse is determined by the parameter gamma(γ), called the spatial aspect Previously the Gray Level Co-occurrence ratio. It has been found to vary in a limited range ofMatrix(GLCM) features are used for classification of 0.23 < gamma(γ) < 0.92. The value of the γ=0.5scene images in two categories natural and artificial . is used., Sigma cannot be controlled directly in theBut we are trying to classify the scene images in two applet. Its value is determined by the choice of thefour different classes using GLCM for features parameters lambda a and b.extraction and PNN as classifier. Our proposed λ & b : The parameter lambda(λ) is themethod is the combination of both Gabor Filter and wavelength and 1 / lambda the spatial frequency ofGLCM .Then the classification accuracy is compared the cosine factor in Eq. (1). The ratio sigma / lambdafor the two methods. determines the spatial frequency bandwidth of simple cells and thus the number of parallelII. GLCM excitatory and inhibitory stripe zones which can be This is a popular statistical texture observed in their receptive fields. The half-responseclassification Technique ever since it is introduced spatial frequency bandwidth b (in octaves) and theby Haralick et al. back in 1973 [11] because it is ratio sigma / lambda are related as follows:computationally simple yet useful for many textureclassification problems. The GLCM calculates theoccurrence of pixel pairs within the imagesaccording to the spatial distance between the pixelpairs and orientations provided [12]. The computedGLCM can be used as a feature after it is down-sampled which we named as raw GLCM in thispaper [9]. A second-order feature can be obtained Neurophysiological research has shown that the half-from the GLCM. There are 5 commonly used response spatial-frequency bandwidths of simpletextural features, i.e. contrast, correlation, energy, cells vary in the range of 0.5 to 2.5 octaves in the catentropy and homogeneity [12]. In this paper, the (weighted mean 1.32 octaves) and 0.4 to 2.6 octavesGLCMs are generated as in [5] and [8]. In this study in the macaque monkey (median 1.4 octaves). Whilewe use the following three features that are most there is a considerable spread, the bulk of cells havecommonly used: bandwidths in the range 1.0-1.8 octaves. Some researchers propose that this spread is due to the Energy  {P(i, j )}2 gradual sharpening of the orientation and spatial i, j frequency bandwidth at consecutive stages of the Inertia   (i  j )2 * P(i, j ) visual system and that the input to higher processing i, j stages is provided by the more narrowly tuned Entropy   P(i, j ) *log( P(i, j )) simple cells with half-response spatial frequency i, j bandwidth of approximately one octave. SinceIII. GABOR FILTER lambda and sigma are not independent when theThis is a signal processing method, therefore it bandwidth is fixed, only one of them, lambda, isProcesses on the frequency domain rather than the considered as a free parameter which is used in thespatial domain [2]. In this paper, the Gabor filters applet.are generated by using different two radial centre Θ: The angle parameter theta specifies thefrequencies and eight orientations as used in [7]. The orientation of the normal to the parallel excitatoryfollowing family of two-dimensional Gabor and inhibitory stripe zones - this normal is the axis xfunctions was proposed by Daugman [1] to model in Eq. (1) - which can be observed in the receptivethe spatial summation properties (of the receptive fields of simple cells.fields) of simple cells in the visual cortex: Φ: Finally, the parameter phi, which is a phase offset in the argument of the cosine factor in Eq. (1), determines the symmetry of the concerned Gabor function: for phi=0 degrees and phi=180 degrees the function is symmetric, or even; for phi= -90 degrees and phi= 90 degrees, the function is antisymmetric, or odd, and all other cases are asymmetric mixtures where the arguments x and y specify the of these two.position of a light impulse in the visual field and The values of Φ used in the simulation are Φ=0(ξ,η) has the same domain as (x,y),σ, γ,λ,θ and Φ for symmetric receptive fields and Φ= -Π/2 forare parameters as follows : antisymmetric receptive fields. Due to theσ & γ : The standard deviation sigma(σ) of the complexity of the features produced, the GaborGaussian factor determines the (linear) size of the filters are down-sampled and the singular value 238 | P a g e
  3. 3. Ranjita Mishra / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 3, Issue 2, March -April 2013, pp.237-241decomposition (SVD) is further used to reduce the VII. EXPERIMENTAL RESULTSdimensionality of the feature set [7]. TABLE I :GABOR ENERGY FOR DIFFERENTIV. COMBINED GLCM AND GABOR FILTER SCENES IMAGE Different techniques are often combined to near near Far Farbe used and can produce better results compared to natural manmade natural manmadeusing them individually [7, 8, 10, 13]. In this paper, 37.5 18.8 16.8 31.4we combine the Gabor filters and the GLCM feature 35.2 17.3 12.5 56.3by appending both of them into a single feature set.This combination produces a better result than either 30.3 21.1 13 45.3GLCM feature or Gabor filters . 30.6 19.9 14 37.7V. PNN 41.4 20 16.9 35 Probabilistic Neural Network can be usedfor classification problems. When an input is TABLE III: CLASSIFICATION ACCURACY-BYpresented , the first layer computes the distance TAKING 12from the input vector to the training input vectors IMAGES FROM EACH CATEGORYand produces a vector whose element indicatehow close the input is to a training input. The second Image Used Classification Accuracy (%)layer sums these contributions for each class of GLCM Gabor & GLCMinputs to produce as its net output a vector of Near Natural 83.3 83.3probabilities . Finally a complete transfer function Near 66.7 78on the output of the second layer chooses the manmademaximum of these probabilities and assigns „1‟ Far_ Natural 75 91to that class and „0‟ otherwise. For example if Far manmade 66.7 83.3there are Q input vector/target pairs and eachtarget vector has K elements, then one of theseelement is 1 and the rest are 0. Thus each input TABLE IIIII:OVERALL CLASSIFICATIONvector is associated with one of the K class. ACCURACYVI. EXPERIMENTAL DATASET Different No of No of % Correct The dataset used in this paper is the real Method images images resultworld dataset and the experimental tool used is Test missMATLAB. Images of manmade objects & natural classifyscenes download from websites[17,18] are GLCM 100 22 78considered for the present study. The algorithmhas been verified on images having two Gabor+GLCM 100 12 88different depth . VIII. DISCUSSION We have created our own database of Scenes of 4 classes each containing 100 scene images. We conducted experimentation under varying database size and we studied its effect on classification accuracy. There are three experiments conducted in this paper. They are conducted on the GLCM, Gabor filters, combined GLCM and Gabor filters, for scene image classification. Gabor FilterSample images belonging to four classes: near- can classify these scene images into two groups lownatural scenes (Row-1), near-manmade objects energy & high energy, but it cannot classify in two(Row-2), Far-natural scenes (Row-3) & Far- four different class.So we compair GLCM and themanmade structure (Row-4) combined of Gabor & GLCM. The experimental results show that the Success rate is of 78% using GLCM & Combined rate of 88% by using both Gabor filter & GLCM.The comparison is shown in Table III. 239 | P a g e
  4. 4. Ranjita Mishra / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 3, Issue 2, March -April 2013, pp.237-241IX. CONCLUSIONS Applications 2007 , IEEE DOI In this paper we have proposed a 10.1109/ICCIMA.2007.77probabilistic neural network based Scene image [3] R. M. Haralick, K. Shanmugam, andclassification method by the use of texture features. I.Dinstein, “Textural Features forThe suitable texture features such as GLCM and ImageClassification”, IEEE Trans. onGabor response is explored for the purpose of Scene Systems, Man, and Cybernetics, Vol. SMC-image classification . We have created our own 3, No.6, November1973, pp.610-621database of Scenes of four classes each [4] Devendran V “Feature Selection for Scenecontaining 100 scene images. We conducted Categorization using Support Vectorexperimentation under varying database size and we Machines” 2008 IEEE Congress on Imagestudied its effect on classification accuracy. and Signal Processing, DOI 10.1109/CISP.2008.579 An alternative to objects for scene [5] J. Garcia-Consuegra, G. Cisnerosrepresentation is to directly compute space “Integration of Gabor Functions with Co-descriptors; the space descriptors (naturalness, occurrence Matrices: Application toopenness, etc.) provide a scene representation at Woody Crop Location in Remote Sensing”multiple levels of categorization and independently 0-7803-5467-2/99/ 1999 IEEEof the complexity of the scene. The descriptors of [6] Shuo Chen, Chengdong Wu, Dongyuethe spatial envelope are correlated with the image Chen, Wenjun Tan “Scene Classificationsecond-order statistics and the spatial arrangement of Based on Gray Level-Gradient Co-structures in the scene. Occurrence Matrix in the Neighborhood of Interest Points” National Nature Science Result of current model: Success rate is of Foundation 978-1-4244-4738-1/09/200978% using GLCM and Combined rate of 88% by IEEEusing both Gabor filter & GLCM. Here, we [7] Zhihua xie , Guodong Liu , Cuiqun He,discovered that the overall accuracy rate produced Yujie Wen “Texture Image Retrievalusing feature images generated by both the Gabor Based on gray level cooccurrence matrixfilters & GLCM has the best accuracy at 88 % for and singular value decomposition” 978-1-the scene image classification problem. And the 4244-7874-3/10/2010, IEEEclassification accuracy, by taking 12 images from [8] Peter Kruizinga and Nikolay Petkoveach category is maximum (83.3%) for near natural N(1999) “Nonlinear Operator for Oriented& far natural by using GLCM, and maximum (91%) Texture ” IEEE Transation,vol8,No10for far natural object by using the combination of pp1395-1407both Gabor filter and GLCM. [9] Oliva A,Torralba A (2002)”DepthThe experimental results have shown that: estimation from image structure “, IEEE• The Gabor filter can classify the scene Transaction on Pattern Analysis andimages into two group based on their energy but the Machine Intelligence , vol.24 , No.9,pp-filter gives very poor result if used for classification 1226-1283into four classes because the energy of the near [10] JOHN G. DAUGMAN “Completenatural and far manmade are very close to each Discrete 2-D Gabor Transforms by Neuralother. Similarly the energy of the near manmade and Networks for Image Analysis andfar natural have the close values. Compression “ IEEE TRANSACTIONS• The GLCM method produces a poor result ON ACOUSTICS, SPEECH, ANDas compared to combination of Gabor filter and SIGNAL PROCESSING. VOL. 36. NO. 7.GLCM classification techniques; JULY 1988 Other Journal Papers:REFERENCES [11] Mari Partio, Bogdan Cramariuc, MoncefIEEE Papers: Gabbouj, and Ari Visa” ROCK TEXTURE [1] Jing Yi Tou, Yong Haur Tay, Phooi Yee RETRIEVAL USING GRAY LEVEL CO- Lau “A Comparative Study for Texture OCCURRENCE MATRIX ” Classification Techniques on Wood [12] Jing Yi Tou1, Yong Haur Tay1, Phooi Yee Species Recognition Problem” 2009 Fifth Lau2 “RECENT TRENDS IN TEXTURE International Conference on Natural CLASSIFICATION: A REVIEW ” Computation, IEEE DOI Symposium on Progress in Information & 10.1109/ICNC.2009.594 Communication Technology 2009 [2] Devendran V , Hemalatha Thiagarajan , [13] N. Petkov, P. Kruizinga” Computational A. K. Santra” Scene Categorization using models of visual neurons specialised in the Invariant Moments and Neural Network detection of periodic and aperiodic “International Conference on oriented visual stimuli: bar and grating Computational Intelligence and Multimedia cells ” Biol. Cybern. 76, 83–96 (1997) 240 | P a g e
  5. 5. Ranjita Mishra / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 3, Issue 2, March -April 2013, pp.237-241 [14] Brandt Tsoa* and Richard C. Olsenb “Scene classification using combined spectral, textural and contextual information “ [15] Julia Vogel and Bernt Schiele, “A semantic typicality measure for natural scene categorization,” Pattern Recognition Symposium DAGM, 2004.[16] Oliva A. and Torralba A., “Modelling the shape of the scene: A holistic representation of the spatial envelope,” International Journal of Computer Vision, vol. 42(3), pp. 145–175, 2001.[17] h/softlib/coil-100.html/[18] 241 | P a g e