Content Based Image Retrieval using Combined Features of Color and Texture Features with SVM Classification

Transcription

1 Content Based Image Retrieval using Combined Features of Color and Texture Features with SVM Classification R. Usha [1] K. Perumal [2] Research Scholar [1] Associate Professor [2] Madurai Kamaraj University, Madurai. Abstract Retrieval of an image is a more effective and efficient for managing extensive image database. Content Based Image Retrieval (CBIR) is a one of the image retrieval technique which uses user visual features of an image such as color, shape, and texture features etc. It permits the end user to give a query image in order to retrieve the stored images in database according to their similarity to the query image. In this work, content based image retrieval is accomplished by combining the two features such as color and texture. Color features are extracted by using hsv histogram, color correlogram and color moment values. Texture features are extracted by Segmentation based Fractal Texture Analysis (SFTA). The combined features which are made up of 32 histogram values,64 color correlogram values, 6 color moment values and 48 texture features are extracted to both query and database images. The extracted feature vector of the query image is compared with extracted feature vectors of the database images to obtain the similar images. The main objective this work is classification of image using SVM algorithm. Keywords Image Retrieval; Content based image retrieval; HSV color histogram; color correlogram; color moments; SVM Algorithm; Relative Standard Derivation; Fractal Texture features. 1. Introduction Nowadays in digital photography, to save bulk of large amounts of high quality images, network speed and storage capacity has been made possible. Digital images are used in a wide range of applications such as geography, medical, architecture, advertising, design, military and albums. However here we have some difficulties in searching and organizing the largest quantity of images in databases. Generally the retrieval of image is classified into two methods such as 1. Text Based Image Retrieval and 2. Content Based Image Retrieval. Text Based Image Retrieval is having following disadvantages such as inefficiency, loss of information, time consuming process and more expensive task. These problems are overcome by using Content Based Image Retrieval for image retrieval. Content based refers that the search will analyse the contents of an image rather than the data about image such as keywords, tags, name of file extension like jpg, bmp, gif etc. Here the content refers visual informations such as color, texture and shape that can be derived from the image itself. Therefore, in this paper we proposed effective CBIR system using color and texture feature to overcome these above mentioned drawbacks of Text based image Retrieval. 2. Related works Image retrieval in CBIR based on the visual features such as texture, color and shape. In this work we choose two visual features as texture and color. Texture analysis, is generally a very time-consuming process. Research in texture analysis is very important, because that is used to improve the discriminatory ability of the extracted image features. There are three primary issues in texture analysis, such as texture classification, texture segmentation and shape recovery from texture. Texture classification, is a process of identifying the given texture region from a given set of texture classes. Texture segmentation is concerned with automatically determining the boundaries between various textured regions in an image [1]. In order to accurately capture the textural characteristics of an image, texture analysis algorithms use filter banks or co-occurrence gray level matrices (GLCMs) have to consider multiple orientations and scales. The computational cost overhead for applying this method may be heavy. It is also reported in [2] that SFTA works much faster in terms of feature extraction time, when compared to Gabor and Haralick methods. The main objective of the SFTA is to extract texture feature in an image which results in the formation of a feature vector. Haussdorf fractal dimension method is used in SFTA. To find optimal threshold Otsu algorithm is used. It is suggested in [3] 169

2 that fractal dimension can be efficiently computed in linear time by the box counting algorithm. For real world images it is suggested in [4], that the Otsu s method provides a better selection of thresholds. In [5], it is reported that Otsu s method the image is assumed to be composed of only two regions: object and background, and the best threshold is the one that maximize the between-classes variance of the two regions. The Otsu s method is also extendable to multilevel thresholding. The color of an image is represented from the famous color spaces like RGB, XYZ, YIQ, L*a*b, U*V*W, YUV and HSV [6]. It has been reported that the HSV color space gives the best color histogram feature, among the different color spaces [7]-[11].In general, histogram-based retrievals in HSV color space showed better performance than in RGB color space. In a viewpoint of computation time and retrieval effectiveness, using HSV color space is faster than using RGB color space [12]. Therefore, in this work we use SFTA texture and HSV color model features are used for efficient image retrieval to improve the above mentioned text based image retrieval problems. 3. Methodology Effective image processing techniques are required to extract visual features such as texture and color from an image in CBIR system. This system accepts the input query image from the user. A retrieval model CBIR performs the image retrieval by comparing the similarities between the input query image and database stored images using these extracted texture and color features. Then the outcome of this system is to find out relevant image from image database to query image given by the end user. The below figure describes the basic functionality of content based image retrieval. Color and texture features are extracted to both query image and database images. Then compare the similarity between the feature vectors of query and database image. Precision and recall operation is carried out for analysis the performance of the system. Figure 1: Block diagram for Content Based Image Retrieval 4. Feature extraction The feature has been defined as a method of one or more measurements, each of which identifies some quantifiable properties of an object, and is calculated such that it quantifies some significant characteristics of the object. Feature extraction is a special form of dimensionality reduction. In this work, an extraction of features consists of color and texture digital information. Color and texture features are extracted by hsv histogram, color correlogram, color moments and SFTA respectively. 4.1 HSV Color histogram Color feature is one of the most important things to access the image. The color of an image is represented from the famous color spaces like RGB, XYZ, YIQ, L*a*b, U*V*W, YUV and HSV [1]. HSV color space gives the best color histogram feature, among the different color spaces [1]. HSV color space is represented by three components such as Hue (H), Saturation(S), and Value (V). A Color histogram of an input image is defined as a following vector H= {H[0], H [1], H[i],, H[N]} 170

3 R Usha et al, International Journal of Computer Science & Communication Networks,Vol 4(5), Where i denotes the color bin in the color histogram. N denotes the total number of bins used in color histogram and H[i] denotes the total number of pixel of color I in an image. Generally, every pixel in an image will be in favour to color histogram bin. So that, in the image color histogram, each bin value gives the number of pixels those have the same corresponding color. Color histogram should be normalized to compare the images in various sizes. The normalized color histogram Hʹ is defined as, Hʹ={Hʹ[0],Hʹ[1],,Hʹ[i],,Hʹ[n]} Where Hʹ[i]= H[i] / Total number of pixels in an image. Algorithm The computation of HSV color histogram has been done by using following steps as, Step1: Convert RGB color image into HSV color space. Step 2: Color quantization is carried out using color histogram by assigning 8 levels to hue, 2 levels to saturation and 2 levels to value for give a quantized HSV space with 8x2x2=32 histogram bins. Step3: The normalized histogram is obtained by dividing with the total number of pixels. 4.2 Color correlogram and color moments Color correlogram gives the information about the features of colors. It includes spatial color correlations, which describes the global distribution of local spatial correlation of colors and is very easy to compute. Color moment feature is used to differentiate images based on their color features and it is also gives the similarity of color measurement between the images. Then the similarity values are compared with the values of image indexed at image database for image retrieval. 4.3 Texture feature Texture features is required here for the below reasons such as Each class has set of images these color is independent but these texture information is dependent from one to one is shown in Figure 2. Images are having same color with different texture is shown in Figure 3. Extraction of texture features may give time consuming process. Solve this time consuming problem by implementing SFTA algorithm [2]. Figure 2 Figure3 An enhanced input RGB image is converted into Grayscale image I. SFTA texture method applied multilevel Otsu thresholding on Grayscale image I for decomposing the segmented image in several parts. This is achieved by selecting pairs of thresholds (lower threshold tl and upper threshold tu) using Two Threshold Binary Decomposition (TTBD). SFTA feature vector correlate with the number of binary images acquired in TTBD phase. If the standard total number of extracted threshold is 4, we acquire 8 different binary images. Each binary image has three feature vectors that depict the boundaries fractal dimension. The purpose of fractal measurement is used to narrate the boundaries complexity and segmented image structures. An extracted vector features are fractal dimension, mean gray level, and size of area image. SFTA algorithm has been explained given below in figure 4. Require: Grayscale image I and number of threshold nt. Ensure: Feature vector VSFTA. 1: T MultiLevelOtsu (I, nt) 2: TA {{ti, ti+1}: ti, ti+1 T, i [1.. T -1]} 3: TB {{ti, nl}: ti T, i [1... ]} 171

4 R Usha et al, International Journal of Computer Science & Communication Networks,Vol 4(5), : i 0 5: for {{tl, tu}: {tl, tu} TA TB} do 6: Ib TwoThresholdSegmentation (I,tl,tu) 7: Δ(x, y) FindBorders (Ib) 8: VSFTA[i] BoxCounting (Δ) 9: VSFTA[i+1] MeanGrayLevel(I, Ib) 10: VSFTA[i+2] PixelCount(Ib) 11: i i+3 12: end for 13: return VSFTA Figure 4: SFTA Algorithm The symbol I, Ib, Δ, T, nt, tl, tu and VSFTA denotes input Grayscale image, binary image, border image, set of threshold values, and total number of thresholds, lower threshold, upper threshold and extracted SFTA feature vectors respectively. 5. Similarity comparison Compare the similarity between the database and query image by using the relative standard deviation. The following equation is defined the Relative Standard Deviation as, SD= 1 N Σ (Xi-X) 2 RSD=stdev/mean* Support Vector Machine Algorithm Support vector machine also known as SVM and is a supervised machine learning method that examine the data and identify the patterns, used for classification. The advantage of this algorithm is to classify the input query object depends on feature vectors and training samples. 7. Performance measurement Generally performance of the CBIR is analysed by calculating the values of precision and recall values. Precision: Precision= Total number of Retrieval Relevant image Total number of Retrieval image Recall: Recall=Total number of Retrieval Relevant image Total number of relevant image 8. Experimental result This proposed approach, the image database contains 400 images. In database images, 100 images are used for testing and remaining 350 images are used for training. An input query image and total number of returned images are desired by the user. Features for query image are extracted by using SFTA and hsv color models. Figure 5: The results of binary image that generate from Two Thresholding Binary Image. There are 16 images output of a single input RGB image 172

5 SVM algorithm classifies the extracted query image features with relevant features of database images. Then calculate the recall and precision value for measuring the performance. Then the extracted feature of image database (Training set) is loaded successfully and it shown below as, 9. Conclusion CBIR is a process to search the relevant image in database image when new or query image is given by the user. In this paper, we use combined color and texture features. Color features are extracted by using hsv histogram; color correlogram, color moments and texture features are extracted by using SFTA. Combined features are extracted to both query and database images (training samples).to classify the query image feature vector with training samples using SVM algorithm and standard deviation is used here for similarity measurement. Performance measurement is calculated by using precision and recall operations. 10. References [1] MihranTuceryan and Anil K.Jain, Texture Analysis, The Handbook of Pattern Recognition and Computer Vision, pp , [2] AlceuFerraz Costa, Gabriel Humpire-Mamani, AgmaJuci Machado Traina, An Efficient Algorithm for Fractal Analysis of texture, Graphics, Patterns and Images (SIBGRAPI), th SIBGRAPI Conference, pp ,

6 [3] C. Traina Jr., A. J. M. Traina, L. Wu, and C. Faloutsos, Fast feature selection using fractal dimension, in Brazilian Symposium on Databases (SBBD), João Pessoa, Brazil, 2000, pp [4] P. K. Sahoo, S. Soltani, A. K. C. Wong, and Y. Chen, A survey of thresholding techniques, Computer Vision Graphics Image Processing, Vol. 41, 1988, pp [5] N. Otsu. A threshold selection method from graylevel histogram. IEEE Trans. Systems Man Cybern. vol. 9, no. 1, 1979, pp [6] T. Gevers, Color in image Database, Intelligent Sensory Information Systems, University of Amsterdam, the Netherlands [7] X. Wan and C. C. Kuo, Color distribution analysis and quantization for image retrieval, In SPIE Storage and Retrieval for Image and Video Databases IV, Vol. SPIE 2670, pp [8] M. W. Ying and Z. HongJiang, Benchmarking of image feature for content-based retrieval, IEEE.Pp , [9] Z. Zhenhua, L. Wenhui and L. Bo, An Improving Technique of Color Histogram in Segmentationbased Image Retrieval, 2009 Fifth International Conference on Information Assurance and Security,IEEE, pp , [10] E. Mathias, Comparing the influence of color spaces and metrics in content-based image retrieval,ieee, pp , [11] S. Manimala and K. Hemachandran, Performance analysis of Color Spaces in Image Retrieval, Assam University Journal of science & Technology, Vol. 7 Number II , [12] Sangoh Jeong, Histogram-Based Color Image Retrieval, Psych221/EE362 Project Report Mar.15,