Model Based Approach for Content Based Image Retrievals Based on Fusion and Relevancy Methodology

Transcription

1 The International Arab Journal of Information Tehnology, Vol. 12, No. 6, November Model Based Approah for Content Based Image Retrievals Based on Fusion and Relevany Methodology Telu Venkata Madhusudhanarao 1, Sanaboina Pallam Setty 2, and Yarramalle Srinivas 3 1 Department of Computer Siene and Engineering, TPIST, India 2 Department of Computer Siene and Systems Engineering, Andhra University, India 3 Department of Information Tehnology, GITAM University, India Abstrat: This paper proposes a methodology for Content Based Image Retrievals (CBIR) using the onept of fusion and relevany mehanism based on KL divergene assoiated with generalized gamma distribution to integrate the features orresponding to multiple modalities, feature level fusion tehnique is onsidered. The relevany approah onsidered bridges the link to both high level and low level features. The target in the CBIR is to retrieve the images of relevany based on the query and retrieving the most relevant images optimizing the time omplexity. A generalized gamma distribution is onsidered in this paper to model the parameters of the query image and basing on the maximum likelihood estimation the generalized gamma distribution, the most relevant images are retrieved. The parameters of the generalized gamma distribution are updated using the EM algorithm. The developed model is tested on the brain images onsidered from brain web data of UCI database. The performane of the model is evaluated using preision and reall. Keywords: CBIR, generalized gamma distribution, relevane image, query image, EM algorithm, preision, reall. Reeived May 23, 13; aepted Otober 12, 13; published online Deember 3, Introdution Content Based Image Retrievals (CBIR) are generally foused on the desription of the low level features of an image. These low level features help to desribe the image, ompare the image and retrieve the relevant images based on the query posed by the user. In pratial situation, sine the usage of internet has drastially improved, retrieving the most appropriate information beomes tedious. In the huge dataset of images retrieved the main hallenge is to identify the idential images by separating the similar images and disarding irrelevant images [14]. CBIR have many appliations ranging from seurity, teleommuniation, Business Proess Outsouring (BPO) [6, 1]. The main objetive of onsidering CBIR in this paper is subjeted to the appliation of its usage in medial domain [3]. With this objetive the methodology is to apply in remote areas for assisting the rural area people and to the dotors with minimum failities to draw onlusions from the available data by the means of ECG, sanning and the other preliminary medial aids. These dataset of the sanned images or the reports available are searhed for the relevany among the available soures to deide the minimum neessary first aids to be adopted to the patient or to literate the patients regarding the disease. Many images an be retrieved based on the searh and some may be relevant and some irrelevant. The relevant images retrieved may not be aurate enough [1, 4, 11, 13]. Hene, to retrieve the images more exatly with relevantly, a methodology to find the relevant images is proposed by using KL divergene algorithm in setion 2. The relevant images obtained are used as the base for the radiologist user to proess the query image and obtain information of relevane [7, 8, 15]. This query based tehnique is presented in setion 3. The output images retrieved may not be legible and hene to have more final details, the feature level fusion tehnique proposed in setion 4 is utilized. The proessed image is sent for the generalized gamma distribution and Probability Density Funtion (PDF) of eah of the images are obtained. Based on the maximum likelihood estimates most relevant images are retrieved. The methodology is presented in setion 5 and the results derived thereof are evaluated using performane metris like preision and reall and are presented in setion Relevant Images Identifiation using KL Divergent Method Majority of the similar images based on the riteria will be retrieved of whih some may be relevant, similar and idential. The ideology is to extrat the most relevant images for whih the PDF [9, 16] of eah of the retrieved image is alulated and for the query image also the PDF s are alulated. In order to find the relevany KL Divergene algorithm [5] is used whih is given by:

2 5 The International Arab Journal of Information Tehnology, Vol. 12, No. 6, November 15 p 1 ( x ) K L ( p1, p 2 ) = p1 ( x ) log dx p 2 ( x ) Where p 1, p 2 are the two PDF omputed on different brain images, formulated using generalized gamma distribution. The set of images retrieved and the images retrieved based on relevany are exhibited in Figures 1, 2-a and b. (1) The images retrieved by proessing the query are shown in Figure Fusion Figure 3. Query images. Fusion addresses the proess of ombining the relevant images from the set of images to get a larity image. Many CBIR tehniques are available in literature [2, 12, 14] and we have onsidered feature level fusion where the most relevant features onsidered in the query image are fused into a single image. The only restrition onsidered for fusion is that the images to be fused should be of same dimension. The output of the fusion image is shown in Figures 4-a, b and 5. a) Images without fusion. Figure 1. Image dataset. b) Images with fusion. Figure 4. Showing Images without and with fusion. a) Relevant Images. b) Retrieved images Figure 2. Showing relevant and retrieved images 3. Query Image Using the query given in Equation 2 one an retrieve an image of interest from the idential set of images [12]. 1 1 Q= αq + β i N = 1 N R γ i N = 1 N N N R N N Where α, β, γ are randomly hosen values with riteria that α+β+γ=1 and N N is the set of non-relevant images. (2) Figure 5. Output after applying fusion. 5. Generalized Gammaa Distribution The generalized gamma distribution is utilized for the purpose of identifying the most similar images based on maximizing the likelihood estimate. Generalized gamma distribution is onsidered beause of its asymmetry and in general the shapes of the body organs are asymmetri in nature [4]. The parameters of the generalized gamma distribution are updated using EM algorithm and are presented below. The PDF of the generalized gamma distribution is of the form:

3 Model Based Approah for Content Based Image Retrievals Based on Fusion 521 (,,,, ) f x k a b x a k 1 b ( x a ) e = k b Γ ( k ) Where a, b and x are alled gamma variants and and k are alled shape parameters. By varying the value of the shape parameters, the partiular ases of gamma distribution an be modeled. The updated Equations 4, 5 and 6 of the generalized gamma distribution using EM algorithm are: ( l+ 1) 1 = 1 f x a ( x a) klog ı + f b x a b log b t k 1 e ( lo g t ) t d t ( l + 1) e k = 1 + x a 1 f Γ ( k 1 ) lo g b f k ( l + 1) b = 6. Methodology k 1 f ( x a ) + 1 b f b The proposed methodology is for the usage of experts deisions for ratifying deease and to suggest the minimum neessary steps to the dotors available in small hospitals at remote areas of rural villages. So, that the patient an be supported with life saving drugs till he is shifted to nearest speialized hospitals. This methodology proposed helps to plan for effetive treatment to the patient. In order to demonstrate the methodology, brain images obtained from UCI medial database is onsidered. The main intension is to identify the type of brain disorders whih inlude Parkinson s disease, Alzheimer s, femur et., In order to retrieve the relevant information, the query image (Brain Sanned MRI Image) of a patient available at the primary health enter is onsidered as the query image. In order to retrieve the relevant images, KL Divergene algorithm proposed in setion 3 is utilized. To have a better quality image or to identify the features inside the Brain images more appropriately the fusion tehnique proposed in setion 4 is utilized. The proessed image is given as input to the PDF of the generalized gamma distribution and basing on the MLE the most relevant images are retrieved. The outputs of the derived models are evaluated using the benh mark metris Preision and Reall are given in Equations 7, 8 and 9 respetively. A Preision= 1 ( A + C ) Where A: No. of relevant images retrieved, C: No. of irrelevant images retrieved and A+C: Total number of irrelevant+relevant images retrieved. Whereas, Reall is the ratio of the number of relevant images retrieved to the total number of (3) (4) (5) (6) (7) relevant images in the database. It is usually expressed as a perentage. A R eall= 1 ( A + B ) Where A: No. of relevant images retrieved, B: No. of relevant images not retrieved and A+B: The total number of relevant images. The Preision and Reall values are tabulated by varying the relevant images and fixing the non-relevant images are presented below in Table 1. The Preision and Reall values are alulated by applying fusion tehniques, proposed in setion 4 and the values are tabulated by varying relevant images and the results are shown below in Table 2. The orresponding graphial values of Preision and Reall for the values of Tables 1 and 2 are presented in Figures 6, 7, 8 and 9. Table 1. Without fusion tehnique. No. of Relevant Images No. of Non-Relevant Ratio of Relevane to Preision Reall in the Database Images in the Database Non-Relevane Table 2. With fusion tehnique. No. of Relevant Images No. of Non-Relevant Ratio of Relevane to Preision Reall in the Database Images in the Database Non-Relevane Preision 1 6 Figure 6. Variation of preision value with R/NR ratio. The number of relative images retrieved is alulated using the Preision and the values obtained are represented in the form of a graph shown in Figure 6. Reall 1 6 Figure 7. Variation of reall value with R/NR ratio. (8)

4 522 The International Arab Journal of Information Tehnology, Vol. 12, No. 6, November 15 Figure 7 represents the graph showing the Reall auray of relative and non-relative images. Preision Figure 8. Variation of preision value with R/NR ratio onsidering fusion and without fusion. Reall s Figure 9. Variation of reall value with R/ NR ratio onsidering fusion and without fusion. Figures 8 and 9 exhibits the ratio of relative and non-relative images against the Preision and Reall auray by using the onept of fusion and without fusion respetively. 7. Conlusions This paper highlights a novel methodology for ontent based image retrieval that an be very muh useful to retrieve the images based on the query and relevane together with fusion. This methodology is presented with the appliation on patients at primary health enter. The outputs of the results derived are evaluated using metris Preision and Reall and are presented whih show good auray. Referenes [1] Burghouts G. and Geusebroek J., Performane Evaluation of Loal Colour Invariants, Computer Vision and Image Understanding, vol. 113, no. 1, pp , 9. [2] Chahooki M. and Charkari N., Shape Retrieval Based on Manifold Learning by Fusion of Dissimilarity Measures, Image Proessing, IET, vol. 6, no. 4, pp , 12. [3] Datta R., Li J., and Wang J., Content-Based Image Retrieval: Approahes and Trends of the New Age, in Proeedings of the 7 th ACM SIGMM International Workshop on Multimedia Information Retrieval, New York, USA, pp , 5. [4] De Ves E., Benavent X., Ruedin A., Aevedo D., and Seijas L., Wavelet-Based Texture Retrieval Modeling the Magnitudes of Wavelet Detail Coeffiients with a Generalized Gamma Distribution, in Proeedings of the th International Conferene on Pattern Reognition, Istanbul, Turkey, pp , 1. [5] Goldberger J., Gordon S., and Greenspan H., An Effiient Image Similarity Measure Based on Approximations of KL-Divergene Between Two Gaussian Mixtures, in Proeedings of the International Conferene on Computer Vision, Nie, Frane, pp , 3. [6] Grigorova A., De Natale F., Dagli C., and Huang T., Content-Based Image Retrieval by Feature Adaptation and Relevane Feedbak, IEEE Transations on Multimedia, vol. 9, no. 6, pp , 7. [7] Hsu C. and Li C., Relevane Feedbak using Generalized Bayesian Framework with Region- Based Optimization Learning, IEEE Transations on Image Proessing, vol. 14, no. 1, pp , 5. [8] Kherfi M., Ziou D., and Bernardi A., Combining Positive and Negative Examples in Relevane Feedbak for Content-Based Image Retrieval, the Journal of Visual Communiation and Image Representation, vol. 14, no. 4, pp , 3. [9] Malik F. and Baharudin B., The Statistial Quantized Histogram Texture Features Analysis for Image Retrieval based on Median and Laplaian Filters in the DCT Domain, the International Arab Journal of Information Tehnology, vol. 1, no. 6, pp , 13. [1] Marakakis A., Siolas G., Galatsanos N., Likas A., and Stafylopatis A., Relevane Feedbak Approah for Image Retrieval Combining Support Vetor Mahines and Adapted Gaussian Mixture Models, Image Proessing, IET, vol. 5, no. 6, pp , 11. [11] Qian F., Li M., Zhang L., Zhang H., and Zhang B., Gaussian Mixture Model for Relevane Feedbak in Image Retrieval, in Proeedings of IEEE International Conferene on Multimedia and Expo, Lausanne, Switzerland, pp , 2. [12] Rui Y., Huang T., and Chang S., Image Retrieval: Current Tehniques, Promising Diretion and Open Issues, the Journal of Visual Communiation and Image Representation, vol. 1, no. 1, pp , [13] Sivakamasundari G. and Seenivasagam V., Different Relevane Feedbak Tehniques in CBIR: A Survey and Comparative Study, in Proeedings of International Conferene on Computing, Eletronis and Eletrial Tehnologies, Tamil, India, pp , 12. [14] Smeulders A., Worring M., Santini S., Gupta A., and Jain R., Content-Based Image Retrieval at the End of the Early Years, IEEE Transations on Pattern Analysis and Mahine Intelligene, vol. 22, no. 12, pp ,.

5 Model Based Approah for Content Based Image Retrievals Based on Fusion 523 [15] Su Z., Zhang H., Li S., and Ma S., Relevane Feedbak in Content-Based Image Retrieval: Bayesian Framework, Feature Subspaes and Progressive Learning, IEEE Transations on Image Proessing, vol. 12, no. 8, pp , 3. [16] Vasonelos N., Minimum Probability of Error Image Retrieval, IEEE Transations on Signal Proessing, vol. 52, no. 8, pp , 4. Telu Venkata Madhusudhanarao reeived his BTeh degree from JNT University, Kakinada, India, and MTeh degree from JNT University Anantapur, India. Currently, he is working as an Assoiate Professor in the Department of Computer Siene and Engineering at Thandra Paparaya Institute of Siene and Tehnology (TPIST), Bobbili. He is pursuing his PhD in the Department of Computer Siene and Engineering, at JNT University, Kakinada, India. His researh interests inlude image proessing, knowledge disovery and data mining, omputer vision and image analysis. Sanaboina Pallam Setty reeived his PhD degree in omputer siene and systems engineering from Andhra University, Visakhapatnam, India. Currently, he is working as a Professor in the Department of Computer Siene and Systems Engineering at Andhra University, Visakhapatnam, India. He has 21 years of teahing and researh experiene. He has guided 4 students for PhD and guiding 12 sholars for PhD. His urrent researh interests are in the areas of image proessing, omputer vision and image analysis, omputer networks, and modeling and simulation. Yarramalle Srinivas reeived his PhD degree in omputer siene with Speialization in Image Proessing from Aharya Nagarjuna University, Guntur, India. Currently, he is working as a Professor in the Department of Information Tehnology at GITAM University, Visakhapatnam, India. He has 17 years of teahing and researh experiene. He has guided two students for PhD and guiding eight sholars for PhD. His urrent researh interests are in the areas of image proessing, knowledge disovery and data mining, omputer vision and image analysis.