International Journal of Computer Engineering and Applications, Volume XII, Special Issue, August 18, www.ijcea.com ISSN 2321-3469 CONTENT-BASED IMAGE RETRIEVAL USING ZERNIKE MOMENTS AND SURF Priyanka Kumari, Department of Computer Science and Engineering Birla Institute of Technology, Ranchi, India priyanka.kmri1993@gmail.com ABSTRACT -Content Based Image Retrieval (CBIR) has been a standout amongst the most mainstream research areas in computer vision for the past 10 years. Retrieving few visually similar images from a huge database containing millions of images has been a challenging task. The proposed technique utilizes the concept of creating a fingerprint of an image by extracting its global as well as local feature. In other words, an image is uniquely represented by means of its global overall appearance as well as by its intricate local details. For global feature extraction Zernike moments have been used up to order 10 and local details are extracted using Speeded up Robust Feature (SURF). Both Zernike moments and SURF allow speedy computation with added advantage of being invariant to geometric transformations. The present work first filters the database images on the basis of their dissimilarity score with the query image computed using Zernike moments. The filtered images are then further screened on the basis of dissimilarity using SURF values. Finally a global dissimilarity is computed which further is used to rank the database images. For performance evaluation, two standard datasets have been used namely WANG and COIL- 10. The precision and recall values obtained with the proposed technique show values higher than most contemporary image retrieval techniques published in past. Keywords Image Retrieval, CBIR, Speeded up Robust Feature, Zernike Moment (ZM). I.INTRODUCTION Image retrieval is a technique of browsing, searching and retrieving image from a large database of digital image collection. The availability of a large amount of digital image group needs powerful algorithms for image retrieval. Therefore in searching for an object from a large collection of the digital image, it is necessary to develop an appropriate information system to efficiently manage these collections. A number of image searching and image retrieval systems Priyanka Kumari 1
CONTENT-BASED IMAGE RETRIEVAL USING ZERNIKE MOMENTS AND SURF have been proposed. One of the well-known techniques in image retrieval is CBIR. CBIR - Content-based image retrieval (CBIR), also known as query by image content (QBIC).CBIR system extracts image information known as the feature that is used to retrieves relevant image from an image database that best matches with the query image. The shape is one of the vital visual highlights in CBIR. These descriptors fall into two classifications Zernike Moment and SURF. CBIR propose a shape descriptor should be relatively invariant, strong, reduced and simple to determine and coordinate. CBIR [7] extract appropriate low-level such as color, texture, shape or their combinations. SURF- Speeded up robust (SURF) [4] is a patented local feature detector and descriptor. It is partly inspired by the scaleinvariant feature transform (SIFT) descriptor. It can be used for the task such as object recognition, image registration, classification or three-d reconstruction. We have used SURF which is faster than other algorithms. The algorithm uses a descriptor based on certain properties and the complexities are further pointed down. Zernike Moments- Zernike moments [6] are complex number moment which uses a set of complex orthogonal polynomials and are defined over a unit circle. Zernike moment is a global shape descriptor.zm for retrieving the binary and gray level image from recognized image database. The work is organized by- Section 2: Literature Review, Section 3: Proposed Methodology, Section 4: Experimental Results and Section 5: Conclusion. II. LITERATURE REVIEW technique with improved retrieval performance due to the global and local descriptor. In this retrieval technique is tested using the standard MPEG-7 shape database and MPEG -7 trademark database. [2] In this article, the author has focused on the CBIR application in the medical domain. It has been proven that CBIR is the promising approach to realize the task of organizing, searching and indexing a large collection of a medical image. [3]The proposed approach combined the texture feature and shape feature for image retrieval on three different databases. The goal of this article is to extract text and shape feature by using GF and ZM, it reflects achievement of more robust feature then using GF or ZM separately. [4]In this article author shows the combination of novel detection, description, and matching steps. It has been analyzed that SURF gives faster and more precise results for interest point detection-description scheme both in speed and accuracy. [5]The performance of SIFT and SURF on various classification dataset proposes an algorithm for segmenting and retrieving the images. In which the approach uses SURF through the feature locally and gives the novel way of retrieving the images effectively and efficiently. [6] The author has analyzed that Zernike moment feature for retrieving the binary and grey level images. Image retrieval by using ZM shows fast computation for retrieving similar image for binary and gray scale images. III. PROPOSED METHODOLOGY In this section gives the explanation of the proposed method for Content Based Image Retrieval using Zernike Moment and SURF.. [1] In this article author has contribute to the research in this field by proposing an innovation trademark retrieval Figure.1.System Architecture of ZM & SURF The architecture explanation is as follows:- this we will be able to utilize the advantage of both the local feature that has been extracted is then stored as a feature vector in the feature database. Online Phase - The online phase where a user can select any image query from an image database. A set of ZM and SURF feature will be extracted from the query image. This feature will later be compared with the same from the set of the earlier stored feature in the feature database. Similarity Measure: It is used to compute the distance measure between the feature from the query image and each feature vector of the database images. Sorting & Ranking: The distance measure will be used to sort and rank the images inside the database so before a set of retrieval output is showed to the user. The various stages involved in the proposed methodology are described in the figure below. Offline Phase - Offline Phase where ZM and SURF their properties were extracted from an image database. The Priyanka Kumari 2
International Journal of Computer Engineering and Applications, Volume XII, Special Issue, August 18, www.ijcea.com ISSN 2321-3469 A. Zernike Moments Zernike moments have many desirable properties, such as rotation invariance, robustness to noise, expression efficiency. The complex ZM are derived from Zernike polynomials which are a set of complex, orthogonal polynomials defined over the interior of a unit circle x 2 + y 2 = 1. (1) Figure.2.Methodology Stages The details of each of the stages involved in the methodology are described below: 1. PREPROCESSING: In the preprocessing phase the noise present in the images are removed and they are made of uniform size. 2. FEATURE EXTRACTION: We will be using two different for representing each image. One is a global shape feature and the other is a local shape feature. By doing and the global descriptor. ZERNIKE MOMENT: Zernike moments are complex number moment which uses a set of complex orthogonal polynomials and are defined over a unit circle. Zernike moment is a global shape descriptor. SURF: Speeded up robust (SURF) is a patented local feature detector and descriptor. It is partly inspired by the scale-invariant feature transform (SIFT) descriptor. To detect interest points, SURF uses an integer approximation of the determinant of Hessian blob detector, which can be computed with 3 integer operations using a pre computed integral image. Where (x,y) is a Zernike polynomial that forms a complete orthogonal set over the interior of the unit disc of. (x,y)= (x,y)exp(jm (2) The Radial Polynomial (x,y) is defined as- Where, (4) Zernike Moments up to 10th order is used. In total 36 Zernike Moments are used. (3) 3. POST PROCESSING: In post processing we combine the Zernike Moments feature vector with SURF feature vector. The final feature vector is then normalized. 4. MATCHING: The feature vector of the query image is matched with the entire set of feature vectors in the database by using Euclidean distance metric. 5. RANKING: The images in the database are ranked according to their similarity with the query image. The image which is at least distance to the quay image is given the highest ranking. Figure.3. Square-to-Circular image transformation In this work, Zernike polynomial will be calculated at each pixel position given that ZM defined in polar coordinates by using a Square-to-Circular image transformation. Zernike polynomial only needs to be computed once for all pixels mapped to the same circle. Figure shows the schematic of Square-to-Circular image transformation. The image pixels are arranged along concentric square can be mapped to concentric circle. The image coordinate system (x,y) is defined with the origin at the center of the square pixel grid. Priyanka Kumari 3
CONTENT-BASED IMAGE RETRIEVAL USING ZERNIKE MOMENTS AND SURF Here, N=Image Size and =Pixels For a given query image, q, is the set of images retrieved for it, is a subset of shapes which are similar to q in and is the set of shapes in the database which are similar to query image. Precision is a measure of retrieval accuracy, and Recall is the measurement for retrieval of relevant images from database. Figure.4. Visual Representation of Zernike Here we are presenting the fractional order of ZM. So this is a shortest description of ZM. This image shows (25) Zernike polynomial. Here n denote order of moments and m denotes repetitions both are equal to value (5) here. The moment of each size are n are orthogonal to each other n=1, it can be easily seen that both the moment are mutually perpendicular and orthogonal with increase the value of n it will become difficult to see the orthogonal due to increase dimensionality. Dataset Description WANG dataset It contains 1000 total images wherein there are 100 images each of 10 different categories. The classes of an image are Tribal, Buses, Beaches, Historical Buildings, Roses, Dishes, Dinosaurs, Mountains, Elephants, and Horses. The dataset of images containing different daily use objects created at Computer Vision Lab (CVL), ETH Zurich-It contains 265 total images where there are 5 images each of different objects. Link-http://www.vision.ee.ethz.ch/en/datasets/ B. Proposed Work The experiments are conducted on the 1000 images are collected from WANG dataset. Then the proposed SURF and ZM technique is applied to the dataset. The proposed method is evaluated using precision, recall and Fmeasure. This parameter is used when all classes consist of same number of images. Precision Recall (P R) is measured using equation- Precision, Recall, F-measure Figure.5. Precision for ZM and SURF Figure.6. Recall for ZM and SURF Figure.7. F-measure for ZM and SURF Table I Precision and Recall Techniques Precision (%) Recall (%) Retrieved using ZM Retrieved using SURF 79 80 92 54 Priyanka Kumari 4
International Journal of Computer Engineering and Applications, Volume XII, Special Issue, August 18, www.ijcea.com ISSN 2321-3469 Techniques Retrieved using ZM Retrieved using SURF Table II F-measure F-measure 0.73 0.82 The present work uses a combination of local and global shape to represent the shape feature in the most appropriate manner. The idea behind implementing Zernike moments and SURF is that both the techniques are invariant to image rotation, scaling and translation. Together Zernike and SURF help us leave no detail of the images, thus creating a robust pattern of each image. ZM can be used to improve the retrieval which is left for feature work. Similarity Computation: The stage mentioned hereunder is similarity measurement stage for image retrieval technique. Figure.5. Similarity Measure The major goal of the propose retrieval technique is two allow the system to retrieve images with both global and local similarity. In the given stages we can explain dissimilarity values are computed. In this stage first dissimilarity computation is the global feature is active and this stage is to filter irrelevant images and to ensure that only images that are globally similar progress to the next stages. Second stage matching computes the dissimilarity values of the local. An average global dissimilarity value is there computed and fixed as the Threshold value. In which all the images with global dissimilarity value is higher than Threshold value are not further measured.the total dissimilarity value is Compute Total Dissimilarity is, are empirical evidence weight set at 0.2 & 0.8 VI. CONCLUSION V. REFERENCE [1] Agrawal, A.S. Jalal, and R. Tripathi. Trademark image retrieval by integrating shape with texture feature. In International Conference on Information Systems and Computer Networks (ISCON), pages 3033, March 2013. [2] Akbarpour, Sh. "A Review on Content Based Image Retrieval in Medical Diagnosis." Technical and Physical Problems of Engineering 5, no. 15 (2013): 148-153. [3] Fu, Xuezheng, Yong Li, Robert Harrison, and Saeid Belkasim. "Content-based image retrieval using gabor-zernike." In Pattern Recognition, 2006. ICPR 2006. 18th International Conference on, vol. 2, pp. 417-420. IEEE, 2006. [4] Herbert Bay, Tinne Tuytelaars, Luc Vran Gool SURF: Speeded Up Robust Features European Conference on Computer Vision,ECCV 2006: Computer Vision ECCV 2006 pp 404-417 [5] Khan, Nabeel Younus, Brendan McCane, and Geoff Wyvill. "SIFT and SURF performance evaluation against various image deformations on benchmark dataset." In Digital Image Computing Techniques and Applications (DICTA), 2011 International Conference on, pp. 501-506. IEEE, 2011. [6] Hitam, Muhammad Suzuri, Suraya Abu Bakar, Wan Nural Jawahir, and Wan Yussof. "ContentBased Image Retrieval Using Zernike Moments for Binary and Grayscale Images." Moments and Moment Invariants-Theory and Applications, GCSR 1 (2014): 275-288. [7] Walia, Ekta, Anjali Goyal, and Y. S. Brar. "Zernike moments and LDP-weighted patches for contentbased image retrieval." Signal, Image and Video Processing 8, no. 3 (2014): 577-594R. [8] W. M. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain, Content-based image retrieval at the end of the early years, IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 22, no. 12, pp.1349 1380, 2000. Priyanka Kumari 5
CONTENT-BASED IMAGE RETRIEVAL USING ZERNIKE MOMENTS AND SURF [9] R. Datta, D. Joshi, J. Li, and J. Z. Wang, Image retrieval: Ideas, influences, and trends of the new age, ACM Comput. Surv., vol. 40, no. 2, pp. 1 60, 2008. [10] K. Mikolajczyk and C. Schmid, A Performance Evaluation of Local Descriptors, IEEE Transactions on Pattern Analysis and Machine Intelligence, 27 (10), pp. 1615-1630, 2005. [11] J.S. Beis and D.Lowe, Shape Indexing Using approximate Nearest-neighbor Search in High- Dimensional Space, in Proceedings of the 1997 Conference on Computer vision and Pattern Recognition(CVPR 97),p.1000,June,1997. [12] C. Veltkamp and M. Tanase, Content-based image retrieval systems: A survey, Technical Report TR UU- CS-2000-34 (revised version), Department of Computing Science, Utrecht University, October 2002. Priyanka Kumari 6