Edge image description using angular radialpartitioningA. Chalechale, A. Mertins and G. NaghdyAbstract: The authors present a novel approach for image representation based on geometricdistribution of edge pixels. Object segmentation is not needed, therefore the input image mayconsist of several complex objects. For an efficient description of an arbitrary edge image, the edgemap is divided into M £ N angular radial partitions and local features are extracted for thesepartitions. The entire image is then described as a set of spatially distributed invariant featuredescriptors using the magnitude of the Fourier transform. The approach is scale- and rotation-invariant and tolerates small translations and erosions. The extracted features are characterised bytheir compactness and fast extraction/matching time. They exhibit significant improvement inretrieval performance using the average normalised modified retrieval rank (ANMRR) measure.Experimental results, using an image database initiated from a movie, confirm the supremacy of theproposed method.1 IntroductionOwing to an overwhelming increase in multimediainformation in relevant databases, there is an urgent needfor efficient tools to manage, search and retrieve suchinformation. Multimedia storage and retrieval has been thefocus of much research in recent years. The field also affectsother disciplines, such as data compression, security andcommunication. MPEG-7 and CBIR (content-based imageretrieval) are the two most important multimedia appli-cations that have addressed this urgent need. MPEG-7 plansto provide a solution for the problem of efficient (fast) andeffective (correct) retrieval through various multimediamaterials. CBIR aims to facilitate the search in imagedatabases based on the image content rather than textretrieval techniques.In most current content-based image retrieval systems theemphasis is on four cues: colour, texture, shape and objectlayout. MPEG-7 suggests descriptors for colour, texture [1]and visual shape [2]. VisualSeek [3], also uses the objectlayout as an image feature. Although colour, texture andshape are significant features for retrieval purposes, theylose their original ability when the query or the databaseimage has no such attributes, for example, when the queryimage is a rough sketch with only black and white lines [4],or when the aim is to search thousands of black andwhite trademarks without a well defined object contour forlogos similar to a given one in a trademark registrationprocess [5].Rotation and translation invariant properties are crucial inmost recognition tasks and should be considered in thefeatures chosen for image retrieval. The invariant methodscan be categorised into the following two main approaches:.Image alignment, i.e. a transformation is applied to theimage so that the object in the image is placed in apredefined standard position. Furthermore, the approachrelies on the extraction of geometric primitives like extremaof the boundary curvature, bitangents or inflection points.Segmentation of the object is necessary and the approach isnot trivial, especially when there exists more than one objectin the image [6]..Invariant features, i.e. using invariant image character-istics which remain unchanged if the object rotates ormoves. Although this approach has attracted considerableinterest [7], it is still based on geometric primitives.Therefore it suffers from the same shortenings as theimage alignment approach.It is desirable to avoid segmentation preprocessing and tostart directly with the image pixels. One possibility is toemploy invariant moments, such as regular moments orZernike moments [8]. However, for eliminating imagetranslations it is necessary to identify at least one matchingpoint between images. The most common choice is thecentre of mass for calculating central moments. Thusmoments can be considered as a hybrid of alignment andinvariant approaches.The edge points hold considerable information about theimage structure especially in the absence of colour=textureinformation or in images where colour and=or texture arenot the discriminating factors. Furthermore, there areapplications, such as sketch-based image retrieval, whereonly the edge map of the database image is comparable tothe sketched query [4, 9, 10]. A face feature representation,called a line edge map (LEM), is proposed in [11] tointegrate the structural information with spatial informationof a face image by grouping pixels of face edge maps intoline segments. In the edge pixel neighbourhood information(EPNI) method, neighbourhood structure of the edge pixelsis used to make an extended feature vector [4]. The vector isused efficiently for comparing sketched queries witharbitrary images. The semantic power of the method isq IEE, 2004IEE Proceedings online no. 20040332doi: 10.1049/ip-vis:20040332A. Chalechale and G. Naghdy are with the School of Electrical, Computerand Telecommunications Engineering, University of Wollongong,Wollongong, NSW 2522, AustraliaA. Mertins is with the Signal Processing Group, Institute of Physics,University of Oldenburg, 26111 Oldenburg, GermanyPaper received 13th October 2003IEE Proc.-Vis. Image Signal Process., Vol. 151, No. 2, April 2004 93examined in [12]. Although the method is scale- andtranslation-invariant, it does not exhibit rotation invariance.Histograms of edge directions are widely used in computervision and image retrieval for edge image matching. Jam andVailaya propose an edge direction histogram for imageretrieval [13] and employ it for trademark registrationprocess. Shih and Chen [14] also use histograms of edgedirection to describe the shapes of representative objects indifferent trademarks. Yoo et al. [15] apply the samehistogram for shape representation in a new proposedcontent-based retrieval system. An edge distribution functionis defined in [16]. A histogram of edge magnitude withrespect to edge orientation angle is used for lane-departuredetection. The edge histogram descriptor (EHD) is proposedin the MPEG-7 standard [17]. The descriptor can be modifiedfor edge map matching by choosing internal parametersappropriately. The retrieval performance is improved byincorporating semi-global and global histograms to thetraditional local histogram [18].A set of angular radial transform (ART) coefficients isused as a region-based shape descriptor in [2, 17]. Thedescriptor describes the shape of an object in an imageefficiently. The object may not only consist of a