Successfully reported this slideshow.
We use your LinkedIn profile and activity data to personalize ads and to show you more relevant ads. You can change your ad preferences anytime.



Published on

  • Be the first to comment

  • Be the first to like this


  1. 1. V.Harichandana, Dr.Sandeep.V.M / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 2, Issue 5, September- October 2012, pp.1315-1319 Efficient And Robust Shape Signatures For Object Recognition V.HARICHANDANA1 Dr.SANDEEP.V.M2 ECE Dept., JPNCE, Mahabubnagar, Andhra Pradesh1 HOD & Prof.,ECE Dept.,JPNCE, Mahabubnagar, Andhra Pradesh2Abstract- Content Based Image Retrieval (CBIR) This paper aims at addressing the CBIRis an important issue in the computer vision challenges using only the shape signatures. Thiscommunity. Both visual and textual descriptions allows us to neglect the other information about theare employed when the user formulates his object, such as color and texture while retrievingqueries. Shape feature is one of the most the images, enabling us to use silhouette of theimportant visual features. The shape feature is images instead of the images themselves. Theessential as it corresponds to the region of paper proposes various shape signatures that helpinterest in images. Consequently, the shape in achieving better recall rates & precision. Therepresentation is fundamental. The shape paper is arranged as follows: The next sectioncomparisons must be compact and accurate, and provides the necessary background for CBIR.must own properties of invariance to several Section 3 deals with the shape descriptors undergeometric transformations such as translation, use and the performance of the present work arerotation and scaling, though the representation evaluated in section 4.The paper are concluded initself may be variant to rotation. This paper section 5.presents simple, efficient and shape descriptorsfor efficient image mining. The main strength of II. BACKGROUNDthe method is its simplicity. Shape analysis involves several important tasks, starting from image acquisition, reaching toKeywords- Distance mapping, Image mining, shape classification. This section gives an overviewMoment invariance, Object Recognition, Shape of three major tasks of shape analysis problem:Descriptor, Shape Recognition, Shape signature. 1) Shape Description: Characterizes the shape and generates a shape descriptor vector (also calledI. INTRODUCTION feature vector) from a given shape. In recent years, content based image 2) Shape Similarity: Establishes the criteria toretrieval has been studied with more attention as allow objective measures of how much two shapeshuge amounts of image data accumulate in various are similar to each other.fields, e.g., medical images, satellite images, art 3) Shape Recognition: Labels the class to thecollections, commercial images and general input shape.photographs. Image databases are usually very big,and in most cases, the images are indexed only by 2.1 Shape Description:keywords given by a human. Although keywords The problem of shape analysis has beenare the most useful in retrieving images that a user pursued by many authors, thus, resulting in a greatwants, sometimes the keyword approach is not amount of research. Recent review papers [6], [8]sufficient. Instead, Query-by-example or pictorial- as well as books [2], [3] provide a good resource ofquery approaches make the system return similar references. In most of the studies, the terms shapeimages to the example image given by a user. The representation and descriptions are usedexample images can be a photograph, user-painted interchangeably. Since some of the representationexample, or line- drawing sketch. methods are inherently used as shape descriptors, Searching for images using shape features there is no well-defined separation between thehas attracted much attention. Shape representation shape representation and description. However,and description is a difficult task. This is because shape representation and description methods arewhen a 3-D real world object is projected onto a 2- defined in [1] as follows. Shape representationD image plane, one dimension of object result in non-numeric values of the original shape.information is lost. As a result, the shape extracted Shape description refers to the methods that resultfrom the image only partially represents the in a numeric values and is a step subsequent toprojected object. To make the problem even more shape representation. For the sake of simplicity, wecomplex, shape is often corrupted with noise, consider the representation and description togetherdefects, arbitrary distortion and occlusion. There throughout the section and refer them as shapeare many shape representation and description description methods.techniques in the literature. 1315 | P a g e
  2. 2. V.Harichandana, Dr.Sandeep.V.M / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 2, Issue 5, September- October 2012, pp.1315-1319 Shape description methods can be satisfied by the shape representation. With theclassified according to the use of shape boundary reference to [1], different similarity methods arepoints or the interior of the shape: Region based Minkowsky Distance, Hausdorff Distance,methods and Boundary based methods. Bottleneck Distance, Turning Function Distance, Frechet Distance, Nonlinear Elastic Matching1) Region Based Methods: Region based shape Distance, and Reflection Distance.descriptors express pixel distribution within a 2-Dobject region. It describes a complex object 2.3 Shape Recognitionconsisting of multiple disconnected regions as well Shape analysis systems extensively useas a simple object with or without holes. Since it is the methodologies of pattern recognition, whichbased on the regional property of an object, the assigns an unknown sample into a pre-defineddescriptor is insensitive to noise that may be class. With reference to [4], numerous techniquesintroduced inevitably in the process of for pattern recognition can be investigated in foursegmentation. Region based methods classified in general approaches:[3] as follows: Moments, Angular Radial 1. Template Matching,Transformation, Shape Decomposition, Shape 2. Statistical Techniques,Matrices and Vectors, Medial Axis Transform, 3. Structural Techniques,Bounding Regions, Scalar Shape Descriptors. 4. Neural Networks. The above approaches are neither necessarily2) Boundary Based Methods: Boundary based independent nor disjoint from each other.shape description methods exploit only objects Occasionally, a recognition technique in oneboundary information. The shape properties of approach can also be considered to be a member ofobject boundary are crucial to human perception in other approaches.judging shape similarity and recognition. Manyauthors, who study on the human visual perception III. EFFICIENT AND ROBUST SHAPEsystem, agree on the significance of high curvature DESCRIPTORSpoints of the shape boundary in visual perception. Shape based image retrieval primarilyIn the psychological experiments, it is suggested involves three steps: shape descriptor, shapethat corners have high information content and, for similarity measures and shape recognition.the purpose of shape description, corners are used 3.1Shape Descriptor:as points of high curvature. Therefore, the shape There are generally two types of shapeboundary contains more information than the shape representations: one is contour based and other isinterior, in terms of perception. Boundary based region based methods. Contour based method needmethods classified in [3] as follows: Polygon extraction of boundary information which in someApproximation, Scale Space Filtering, Stochastic cases may not available. Region based methods,Representation, Boundary Approximation, Set of however, do not necessary rely on shape boundaryBoundary Points, Fourier Descriptors, Coding, and information, but they do not reflect local featuresSimple Boundary Functions. on shape. So in this experiment for generic purposes, both types of shape representation are2.2. Shape Similarity Measurements necessary. Many pattern matching and recognition 1) Scalar Shape Descriptor: The large number oftechniques are based on a similarity measures scalar shape description techniques is presented bybetween patterns. A similarity measure is a heuristic approaches, which yield acceptable resultsfunction defined on pairs of patterns indicating the for description of simple shapes. A shapedegree of resemblance between the patterns. It is description method generates a shape descriptorpossible that our prior knowledge of objects plays a feature vector from a given shape. The requiredsignificant role in our similarity judgments, a role properties of a shape description scheme arewhich may vary considerably depending on the invariance to translation, scale and rotation. Scalarshapes we view. Since perceptual similarity is not a shape descriptor includes the following featureswell-known phenomenon, none of the available like eccentricity and aspect ratio.similarity measures are fully consistent with theHuman Visual System. 2) Simple Boundary Functions: The following In this section, we list some desirable descriptors are mostly based on geometricproperties of similarity measures. Depending on the properties of the boundary. All of them areapplication, a property, which is useful in some sensitive to image resolution. The following arecases, may be undesirable in some other cases. some of the geometric descriptors like centroidCombinations of properties may be contradictory. distance and circularity.While some of the properties are satisfied by thedistance function and the algorithm used in 3) Shape signature by level set method: The levelsimilarity calculation, the others are inherently set technique is a geometric deformable model 1316 | P a g e
  3. 3. V.Harichandana, Dr.Sandeep.V.M / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 2, Issue 5, September- October 2012, pp.1315-1319implemented to segment a given image to extract three features i.e., eccentricity, circularity andthe region of interest. The output of this technique centric distance for the database. These features foris a distance mapped function wherein the the images in the database are computed andboundary of the object is zero level set and other stored. The features of the query image arepoints are assigned signed distance from the computed and the Euclidian distance to the mean ofboundary of the object segmented by level set the features for each class is then computed. Thetechniques [10], [9], and [11]. query image is classified through Nearest Neighbor The shape signatures are obtained from method. The retrieval rate was found to be poorthe distance mapped level set function. The number i.e.60%. This low retrieval rate alarmed us aboutof points with different distance from the boundary the inadequacy of feature set and 3 more features-can be a good shape signature. Here, number of aspect ratio, centroid distance & distance mappedpixels on the object boundary I0, unit distance away signatures, were added.from boundary I1 and two distances away fromboundary I2 has unique relationship that depends onthe shape of object. The normalize difference arecomputed byCity block distance mapping is more suitable forthis shape signature than Euclidean distances. Thisprovides an additional advantage by reducing thecomputational complexity. Fig1: Query: CUP3.2 Shape similarity The shape descriptors eccentricity,elongatedness, centroid distance, circularity, r10, r20,r21, provides an excellent feature set indiscriminative the shape of different classes. Itprovides a large distance between classes and at thesame time maintains lower distances for objectsbelonging to same class.3.3. Shape recognition Shape of object has a strong connection toimage retrieval, where the task is to retrieve a“matching” image from a (possibly large) database.The best match can then be determined after theobjects present have been recognized.1) Feature analysis and matching technique Fig2: Query: ELEPHANT The simplest way of shape recognition isbased on matching the stored prototypes against the Some sample retrieved images are shown in figuresunknown shape to be recognized. General 1 and 2. Here, in each set, the top image is thespeaking, matching operations determines the query image and top 20 retrieved images in thedegree of similarity between two vectors in the descending order of match are shown. The scalefeature space. The set of features those represents a and rotation invariance of the retrieval can be easilycharacteristic portion of a shape or a group of observed. The mismatched shapes have someshapes is compared to the feature vector of the resemblances to the query image.ideal shape class. The description that matchesmost “closely” according to the distance measure Many different measures for evaluatingprovides recognition. the performance of image retrieval systems have been proposed. The measures require a collection3.4. Performance evolution: images in database and a query image. The Experiment deals with 16 classes of common retrieval performance measures –precisionimages in database each class containing 20 and recall are used to evaluate.images. The experiment started considering only 1317 | P a g e
  4. 4. V.Harichandana, Dr.Sandeep.V.M / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 2, Issue 5, September- October 2012, pp.1315-13191) Precision: Precision is the fraction of the shapes The overall performance of our method isretrieved that are relevant to the users’ measured in terms of Recall rate & Precision. Therequirements. rigorous experiments are conducted to evaluate the performance. The table 1 shows the results of retrieving top 5, top 10, top 15 & top 20 shapes from the database. With all the features discussed,2) Recall: Recall is the fraction of the shapes that the retrieval rate has increased. From the table itare relevant to the query that are successfully can be seen that the monotonous decrease inretrieved. precision with increase in no. of images retrieved indicates that the result are not accidental. TABLE 1 Recall rates & Precision for various query shapes. Relevant Total Recall PrecisionS.No Input Top- Top- Top- Top- Releva nt R(20) P(5) P(10) P(15) P(20) 5 10 15 201 Apple 5 9 13 14 20 70% 100% 90% 87% 70%2 Bat 5 10 13 15 20 75% 100% 100% 87% 75%3 Bottle 5 10 15 17 20 85% 100% 100% 100% 85%4 Car 5 10 15 17 20 85% 100% 100% 100% 85%5 Child 5 8 12 14 20 70% 100% 80% 80% 70%6 Cup 5 10 15 16 20 80% 100% 100% 100% 80%7 Box 5 10 15 20 20 100% 100% 100% 100% 100%8 Flower 5 10 13 13 20 65% 100% 100% 87% 65% Elepha9 5 10 15 17 20 85% 100% 100% 100% 85% nt10 Horse 5 10 13 14 20 70% 100% 100% 87% 70%11 Snail 5 10 12 12 20 60% 100% 100% 80% 60%12 Teddy 5 10 13 15 20 75% 100% 100% 87% 75% Average 76.6 100 97.5 91.5 76.6 IV. CONCLUSION & FUTURE WORK query, then repeating the search with the new Shape is one of the most valuable features identify or describe objects represented in As a major conclusion we stand that ourimages. This paper presents a simple and efficient method demonstrated usefulness and effectivenessmethod based on a few set of image features to for both retrieval and recognition purpose,describe shapes. This method aims to be simple and particularly if taken into account its result in a short description. Several improvements are intended to be References:carried as future work. A first one is to learn [1] Chan Tony F, Vese Lumanita A, “Activefeature weights using, as for instance, evolutionary contours without edge”, IEEE Trans. “Onalgorithms (e.g. genetic algorithms) to properly Image processing”, 10(2): 266-277, 2001.tune the used similarity distance metric. This [2] L. F. Costa and R. M. Cesar Jr. “Shapeprocess is expected to increase the accuracy of the Analysis and Classification: Theory andclassifier for a given dataset. These results can be Practice”. CRC Press, 2001.also valuable for retrieval purposes if these weights [3] R. O. Duda, P. E. Hart, and D. G. Stork.demonstrate stability among several datasets. “Pattern Classification”. John Wiley and Another improvement to the retrieval Sons”, Inc., New York, 2001.process is to make use of relevance feedback, [4] M. Hagedoorn. “Pattern Matching Usingwhere the user progressively refines the search Similarity Measures”, PhD Thesis.Utrechtresults by marking images in the results as University, 2000."relevant", "not relevant", or "neutral" to the search 1318 | P a g e
  5. 5. V.Harichandana, Dr.Sandeep.V.M / International Journal of Engineering Research and Applications (IJERA) ISSN: 2248-9622 Vol. 2, Issue 5, September- October 2012, pp.1315-1319 [5] A.E. Johnson and M. Hebert.”Recognizing Processing, Pattern Recognition, Communication, objects by matching oriented points”, Electromagnetics. He is Reviewer for Pattern Proc. IEEE “Conf. Computer Vision and Recognition Letters (PRL). He acted as Reviewer Pattern Recognition”, pages 684–689, for many International Conferences.He has 24 1997. years of teaching experience. He is member of [6] S. Loncaric. “A survey of shape analysis LMIST – Life Member Instrument Society of India techniques”.” Pattern Recognition”, (IISc, Bangalore). And he guided more than 100 31:983–1001, 1998. projects at UG level and 35 at PG level.And [7] Nafiz Arica.”SHAPE: Representation, published more than 10 papers in international Description, Similarity and Recognition”, journals and 9 Conference Proceedings. PhD Thesis. [8] T. Pavlidis.” A review of algorithms for shape analysis”. “Computer Graphics Image Processing”, 7:243–258, 1978. [9] Sandeep V.M, Subhash Kulkarni, Vinayadatt Kohir,”Level set issues for efficient image segmentation”, International Journal of “Image and Data Fusion”, Vol2, No1, March 2011, 75-92. [10] Sandeep V.M, M, Narayana, Subhash Kulkarni, ”A Scale & rotation invariant fast image mining for shapes’, IEEE conference on “AI tools in Engineering”,Pune,India,2008 [11] Sandeep V.M, Subhash Kulkarni, ”Curve invariant fast distance mapping technique for level sets”, IEEE’s ICSIP 2006, Hubli, India,777-780,2006.Authors Information:1.V.Hari Chandana Pursuing M.Tech(DSCE)from JayaPrakash Narayana College ofEngineering B.Tech(ECE) from JayaPrakashNarayana College of Engineering Currently she isworking as Assistant Professor at Jayaprakashnarayan college of engineering And has 2 years ofExperience in teaching . Her areas of interestinclude,Image Processing ,wireless networks,signal processing.2. Dr. Sandeep V.M. completed Ph.D in Faculty ofElectrical and Electronics Engineering, Sciences,from Visveswaraiah Technological University,Belgaum, and M.Tech from Gulbarga Universityand B.Tech from Gulbarga University. His researchinterests are in the areas of Signal and Image 1319 | P a g e