Published on

IJERA (International journal of Engineering Research and Applications) is International online, ... peer reviewed journal. For more detail or submit your article, please visit www.ijera.com

  • 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. Gurupjit Singh Brar, Amandeep Kaur / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.1742-1747 Analysis Of Rotation Invariant Template Matching Techniques For Trademarks Gurupjit Singh Brar*, Amandeep Kaur** *(Deptt. of Computer Science, Punjabi University, Patiala) ** (Assistant Professor, Deptt. of Computer Science, Punjabi University, Patiala)ABSTRACT As every product has the trademark and their difference. It is the process of moving thefor matching process of that trademark, we need template within the large image at every position ina template matching technique. Further, a which template has to be searched by evaluating thetrademark can be rotated through any angle and similarities between the template and the imagefor this rotation; invariant template matching region over the template’s position and thentechniques are used. For getting the best determining the position based on the similaritytechnique, we are analyzing the ring projection measure.transformation and orientation techniques. In In conventional template matching crossring projection, transformation process the correlation is the most commonly used algorithms.template is converted to a 1D level signal as a The method is simple to implement and understand,function of radius and then the matching is but it is one of the slowest methods. Furthermore, ifperformed. Another process which is used in the the object in the image is rotated, the traditionanalyses is orientation codes. It is based on the method can fail. Thus to solve this type of problem,utilization of the gradient information in the we use rotation invariant template matchingform of orientation codes as the feature for techniques.approximating the rotation angle as well as for Using rotation invariant template matchingmatching. the object of interest can be easily searched even if it is rotated at any angle within the query image.Keywords: Ring projection transformation, Rotation invariant template matching is used inorientation codes, trademarks, rotation invariant. many applications like, trademark detection, character recognition, fingerprint identification,1. INTRODUCTION biomedical imaging, remote sensing and feature In the image processing systems, the input tracking. There are a number of techniques that areimage is first preprocessed, which may involve being used for rotation invariant template matchingrestoration, enhancement, or just proper such as normalized invariant moments [1], Radonrepresentation of the data. After this, some features transform[4], wavelet decomposition [6], ringare extracted for segmentation of the image into its projection transformations [13][14] and orientationcomponents. Then the image classification is done codes [15][16][17]. In this paper we analyze twothat maps different regions or segments into one of techniques i.e. Ring Projection Transformation andthe several objects, each which is identified by a Orientation Codes for rotation invariant trademarklabel. The most significant problem in image detection. These techniques are simple to implementanalysis is the detection of presence of an object and give high accuracy for applications likewithin a given scene or an image. Such problem character recognition.occurs in trademarks classification and other areas In the field of image processing a lot oflike remote sensing for monitoring growth patterns work has been done on template matchingwithin urban areas, weather prediction from techniques that are invariant to rotation. Goshtasbysatellite, target detection from radar or from the A. [1] proposed a rotationally invariant templatefighter plane, etc. For this purpose, template matching using normalized invariant moments. It ismatching or template detection is the most shown that if normalized invariant moments incommonly used technique. circular windows are used, then template matching Template Matching is one of the most in rotated images becomes similar to templatecommon techniques used in Image Processing. It is matching in translated images. Schmidt W. et al. [2]the process of determining the position of sub image considered the higher order neural networks that areinside the large image. The two terms template and robust in terms of target detect ability. Waveletmatching to have their technical meanings: template decomposition approach for rotation invariantis anything that is fashioned, shaped, or designed to template matching is proposed by Du-Ming T. et al.serve as a model from something is to be made, and [6]. They first decompose the image into differentmatching is the process of comparing in respect of multi-resolution levels in the wavelet-transformedsimilarity so that we can examine their likeness or domain, and use only the pixels with high wavelet 1742 | P a g e
  2. 2. Gurupjit Singh Brar, Amandeep Kaur / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.1742-1747coefficients in the decomposed detail sub images to template matching is used in their paper; in the firstcompute the normalized correlation between two stage, histograms of orientation codes are employedcompared patterns. To make the matching invariant for approximating the rotation angle of the objectto rotation, they further use the ring-projection and then in the second stage. Matching is performedtransform. Kim H. Y. et al. [8] considered the by rotating the object template by the estimatedgrayscale template matching problem, invariant to angle. Matching in the second stage is performedrotation, scale, translation, brightness and contrast. only for the positions which have higher similarityThey proposed a technique that substantially results in the first stage, thereby pruning outaccelerates this searching while obtaining the same insignificant locations to speed up the search.results as the original brute force algorithm. Their Experiments with real-world scenes demonstratealgorithm consists of three cascade filters. Xia L. et the rotation and brightness invariance of theal. [9] proposed a learning-based logo recognition proposed method for performing object search.method to detect and classify the logos in the natural Zhonghai L.I. et al. [16] present the concept ofimage. In this they first applied the SIFT matching relative orientation code, and designed a templatesolution in a set of training data to reliably detect the matching method based on relative orientation codeinterest region in images and extract the in order to improve accuracy and real time ability.discriminate features. Second, the approximate This method first selects candidate target pointsnearest neighbor searching strategy is built up by using relative orientation codes histogram matching,formulating the data into a tree-based structure, for then calculates the rotation angle of candidate subefficient matching. Finally, to recognize the logos in image and rotates the sub image, finally match thethe test image, the corresponding SIFT features will relative orientation codes in the sub image. Theirbe computed in the interest regions of test image and results show that this method is robust when thematched in the database which is achieved in the image is rotated by any angle, and the candidatetraining stage. points are selected and matched accurately, and the Further the research has also been done on computing time is less than original method. Uranothe techniques like Ring projection transformation et al. [17] proposed a fast and rotation invariantand orientation codes for the purpose of Rotation template matching using an orientation codeinvariant template matching. Lin Y-H. et al. [13] difference histogram (OCDH). The method isproposed a new method for template matching, effective in presence of some irregularities likeinvariant to image translation, rotation and scaling. shading, highlighting, occlusion or theirIn the first step of the approach, the ring-projection combination. This method is based on Orientationtransform (RPT) process is used to convert the 2D code matching (OCM) and orientation codetemplate in a circular region into a 1D gray-level histogram.signal as a function of radius. Then, the templatematching is performed by constructing a parametric 2. ROTATION INVARIANTtemplate vector (PTV) of the 1D gray-level signal TECHNIQUE IN BRIEFwith differently scaled templates of the object. Lee 2.1 Ring Projection Transformation (RPT)W-C. et al. [14] proposed an algorithm for rotation In order to make our technique rotationinvariant template matching method. The algorithm invariant we use Ring Projection Transformationconsists of two stages process. In the first stage, the Technique [13, 14] which reduces the 2D image intoscene image and template are reduced in resolution. 1D feature vector that is used for matching. ForThen the features from the sub image covered by the applying this process, a template T(x,y) is definedtemplate are extracted using RPT method. The whose size is M×N. The RPT of the template isnormalized correlation formula is used to select the calculated as follows: Firstly, the center point of thematching candidates in the coarse search stage. template T(x,y) denoted as (xc,yc ) is derived, andThese candidates will be recovered to original subsequently, the Cartesian frame template T(x,y) isposition and determine the searching range. In the transformed into a polar frame based on thesecond-stage process, Zernike moments based on following relations:the matching candidates are used to determine theoptimal matching point. Ullah F. et al. [15] proposeda new method for rotation-invariant templatematching in gray scale images. It is based on the where ,utilization of gradient information in the form oforientation codes as the feature for approximatingthe rotation angle as well as for matching.Orientation codes-based matching is robust for The RPT value of the template T(x,y) at radius r issearching objects in cluttered environments even in denoted as PT (r), which is defined asthe cases of illumination fluctuations resulting fromshadowing or highlighting, etc. A two-stageframework for realizing the rotation-invariant 1743 | P a g e
  3. 3. Gurupjit Singh Brar, Amandeep Kaur / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.1742-1747Where Sr is the total number of pixels falling on the estimates the approximate rotation angle of thecircle of radius r. According to the formula of PT (r), object image by using histograms of the orientationit is easily seen that PT (r) is defined as the mean of codes and the second stage searches for the truepixel intensities along the circle centered on the position by performing match only on the candidatestemplate T(x,y). RPT values at the radius r, r∈[0,R], filtered out in the first stage and rotating thecan be regarded as features with equal importance in template by the closest angle of rotation alsothe computation of correlation. The RPT process is estimated in the first stage.constructed along circular rings of increasing radii.The derived one dimensional ring projectiontemplate is invariant to rotation of its correspondingtwo-dimensional image template. Fig. 2 Illustration of orientation Codes [15]Fig. 1 The ring projection of template [13] 3. IMPLEMENTATION 3.1 Ring Projection Transformation (RPT)2.2 Orientation Codes The RPT process is constructed along The orientation codes, for discrete images, circular rings of increasing radii. The derived one-are obtained as quantized values of gradient angle dimensional ring projection template is invariant toaround each pixel by applying differential operator rotation of its corresponding two-dimensional imagefor computing horizontal and vertical derivatives template.like Sobel operator, and then extracting the gradient After calculating the RPTs of the template,angle as follows: we traverse the template on the image pixel by pixel and correspondingly calculating the RPTs of theThe orientation code for a pixel location (i,j), for a region within the image which is traversed by thepreset sector width ∆θ is given as follows: template at that point of time. Then simultaneously compare the RPTs of the template with the RPTs of that region within the image and calculate the difference. After getting all the differences while traversing template on the image, we find the minimum difference value and corresponding to Where [ · ] is the Gaussian operation. If that, we get the match of the template which wethere are N orientation codes then C(ij) is assigned were finding.values {0,1,……,N − 1}. Code N is substituted for Following is the steps that are followed for thislow contrast regions (defined by the threshold Г) for technique:which it is not possible to stably compute the Step 1: Read the input image and the template that isgradient angles. For a separate image O = {Cij} to be found.(referred to as an orientation code image hereafter). Step 2: Calculate the radii at each pixel position onThe threshold Г is important for suppressing the the template. For instance, if we are having theeffects of noise and has to be selected according to template of the size 5x5 pixels, then the radius atthe problem at hand; very large values can cause each pixel of the template will be:suppression of the texture information.Orientation Code Matching (OCM) has proved to berobust for template matching in the presence ofmany real world irregularities, such as illuminationvariation due to shading, highlighting, and in thepresence of partial occlusion. The first stage 1744 | P a g e
  4. 4. Gurupjit Singh Brar, Amandeep Kaur / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.1742-1747 4 3 3 3 4 4. RESULTS Experiments have been performed to test 3 2 2 2 3 the proposed algorithms and to measure the 3 2 1 2 3 accuracy of the algorithms, 52 grayscale images of various trademarks collected are used for testing the 3 2 2 2 3 techniques. All the trademark images have been 4 3 3 3 4 obtained from the internet, and all the trademarks are normalized to 64x64 for fast processing. QueryStep 3: Then the RPT values of the template is images are synthesized by adding a number ofcalculated by taking the sum of the pixel values at trademark images rotated through different angles.the particular radius divided by the number of pixels All the query images are of size 256x256.on that radius. Trademarks are then searched in the query images.Step 4: Traverse the template on the image pixel by Different results are obtained while performing thepixel and correspondingly calculate the RPT values Ring Projection Transformation and Orientationwithin the sub window of the image which is codes technique.traversed by the template at that point of time. The visual results for query image shownStep 5: Compare the RPT values of the template in Fig 3 and template image Fig 4 are shown in Figwith the RPT values of sub window of the image 5 for the RPT method and Fig 6 shows the resultsand calculate the sum of absolute differences. for OC method. It is clear from the visual resultsStep 6: Find the minimum of the sum of absolute that RPT method gives more accurate results thandifference and highlight the corresponding area as OC method for Trademark images where thethe region of the desired template. background is not cluttered and there are no non uniform illumination effects.3.2 Orientation Codes The first stage estimates the approximaterotation angle at the object image by usinghistograms of the orientation codes and the second-stage searches for the true position by performingmatch only on the candidates filtered out in the firststage and rotating the template by the closest angleof rotation also estimated in the first stage.Following are the steps that are followed for thistechnique:Step 1: Read the input image and the template imagethat is to be found.Step 2: Compute the horizontal and verticalderivatives and then compute the gradient angle ofthe template.Step 3: Calculate the orientation codes for all pixellocations and then form the orientation codehistogram of the template.Step 4: Overlap the starting point of the image bytemplate and correspondingly form the orientationcode histogram of the each sub window overlappedby calculating the gradient angle and the orientationcodes of the sub window of the image.Step 5: Shift the histogram of the sub window one Fig. 3 Query Imageby one and correspondingly calculate the absolutedifference between the histograms of the templateand the sub window overlapped by the template.Find the minimum difference between the twohistograms.Step 6: Traverse the template on the image pixel bypixel and repeat the steps 4 and 5.Step 7: Find the minimum difference of the valuesobtained in step 6 and correspond to that find the Fig. 4 Trademarkindex and highlight the corresponding area as theregion of the desired template. 1745 | P a g e
  5. 5. Gurupjit Singh Brar, Amandeep Kaur / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.1742-1747 REFERENCES [1] Ardeshir Goshtasby, “Template Matching in Rotated Images”, IEEE Transactions on pattern analysis and machine intelligence, vol. PAMI-7, No.3, pp. 338-344, May 1985. [2] William A. C. Schmidt and Jon P. Davis, “Pattern Recognition Properties of Various Feature Spaces for Higher Order Neural Networks”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 15. no. 8, August 1993, pp. 795-801. [3] Anil K. Jain Yu Zhong and Sridhar Lakshmanan, “Object Matching Using Deformable Templates”, IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 18, no. 3, 267- 278, March 1996. [4] Osama K. Al-Shaykh and John F. Doherty, “Invariant Image Analysis Based on Radon Transform and SVD”, IEEE TransactionsFig. 5 Resulting Image for RPT image. The on Circuits and Systems-11: Analog anddetected trademark is enclosed in a square. Digital Signal Processing, vol. 43, no. 2, 123-133, February 1996. [5] Goh Wee Leng & D. P. Mital, “A System for Trademark Pattern Registration and Recognition”, Second International Conference on Knowledge-Based Intelligent Electronic Systems, 21-23 April 1998, Adelaide, Australia, pp. 363-369. [6] Tsai Du-Ming, Chiang Cheng-Huei, “Rotation invariant pattern matching using wavelet decomposition”, Pattern Recognition, 2002, 23(13):191-201. [7] Roger M. Dufour, Eric L. Miller and Nikolas P. Galatsanos, “Template Matching Based Object Recognition with Unknown Geometric Parameters”, IEEE Transactions on image processing, vol. 11, no. 12, 1385-1396, December 2002. [8] H. Y. Kim and S. A. Araujo, “Grayscale Template-Matching Invariant to Rotation, Scale, Translation, Brightness and Contrast” , IEEE Pacific-Rim SymposiumFig. 6 Resulting Image for OC image. The on Image and Video Technology, Lecturedetected trademark is enclosed in a square. Notes in Computer Science, vol. 4872,The methods were tested using 52 test images. For pp.100-113, 2007.the RPT the accuracy was 100% and for OC method [9] Goh Wee Leng & D. P. Mital, “A Systemthe accuracy was below 70%. for Trademark Pattern Registration and Recognition”, Second International5. CONCLUSIONS Conference on Knowledge-Based From above experimental results we have Intelligent Electronic Systems, 21-23 Aprilfound that Ring Projection Transformation process 1998, Adelaide, Australia, pp. 363-369.is better than Orientation Codes for matching the [10] M.A. Manzar, T.A. Cheema, A. Jalil androtated trademarks. The results will be further tested I.M. Qureshi, “New image matchingusing standard database of trademark images having technique based on hyper-vectorisation ofnatural images with distortions like noise clutter and grey level sliced binary images”, IETnon uniform illumination effects. Image Process., 2008, Vol. 2, No. 6, pp. 337–351. 1746 | P a g e
  6. 6. Gurupjit Singh Brar, Amandeep Kaur / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 www.ijera.com Vol. 2, Issue 5, September- October 2012, pp.1742-1747[11] S. Xu, “Robust traffic sign shape [16] LI Zhonghai, LIU Chunyu, CUI Jianguo, recognition using geometric matching”, SHEN Weifeng, “Improved Rotation IET Intell. Transp. Syst., 2009, Vol. 3, No. Invariant Template Matching Method 1, pp. 10–18. Using Relative Orientation Codes”,[12] Mahdi Amiri and Hamid R. Rabiee, Proceedings of the 30th Chinese Control “RASIM: A Novel Rotation and Scale Conference July 22-24, 2011, Yantai, Invariant Matching of Local Image Interest China, pp.3119-3123. Points”, IEEE Transactions on Image [17] Takahiro Urano, Shun’ichi Kaneko, Processing, vol. 20, no. 12, 3580-3591, Takayuki Tanaka, Munatoshi Imada, December 2011. “Using Orientation Code Difference[13] Yi-Hsien Lin, Chin-Hsing Chen, Histogram (OCDH) for Robust Rotation- “Template matching using the parametric invariant Search”, Conference on Machine template vector with translation, rotation Vision Application, May 16-17, 2005, and scale invariance”, Pattern Recognition Tsukuba Science City, Japan, pp 364-367. 41 (2008), pp. 2413 – 2421. [18] Anil K. Jain and Aditya Vailaya, “Shape-[14] Wen-Chia Lee, Chin-Hsing Chen, “A Fast Based Retrieval: A Case Study with Template Matching Method for Rotation Trademark Image Databases, Pattern Invariance Using Two Stage Process”, Recognition, 31(9), 1369-1390,1998. IEEE, 2009, pp.9-12. [19] Jan Schietse, John P. Eakins, Remco C.[15] Farhan Ullah, Shun’ichi Kaneko, “Using Veltkamp, “Practice and Challenges in Orientation codes for rotation-invariant Trademark Image Retrieval”, CIVR’07, template matching”, Pattern Recognition July 9-11, 2007. 37(2004), pp. 201-209. 1747 | P a g e