5.30 Image Representation and Recognition Where, yeh p+q=2,3 (5) In addition to all these moments, a set of seven invariant moments can be derived from second and third moments. These are invariant to translation, rotation and scale change. All the above said moments can be used as descriptors for a region of an image. 5.5 OBJECT RECOGNITION - Introduction Recognition is the process that assigns a ‘label’ to an object of interest ifi an image, based on its descriptors. Objects are just the individual image regions. Classification The pattern recognition approaches are categorized into two as (1) Decision-theoretic — This deals with patterns that are described using quantitative descriptors such as length, area and texture. (2) Structural — This deals with patterns that are described by qualitative descriptors such as relational descriptors. 5.6 PATTERNS AND PATTERN CLASSES - A pattern is an arrangement of descriptors. A descriptor is also known as the feature. - A pattern class is a family of patterns that share some com mon properties. Pattern classes are denoted as 0, ©, Where, W is the number of classes, - Pattern recognition is the process of assigning patterns to their respective classes. This is done by using machines and with a little human intervention. Pattern Arrangements Three pattern arrangements that are commonly used in practice are (1) Vectors ~ for quantitative descriptions, (2) Strings ~ for structural descriptions (3) Trees — for structural descriptions Scanned with CamScannerDigital Image Processing 5.31 5.6.1 Vectors Patterns vectors are used for the pattern arrangement of quantitative descriptors such as length, area and structure, These vectors are represented as columns ic n x 1 matrices as shown in the form below. x 1 x, X= i: Street 9 (5.35) Xn Where: T - Transposition, n—Total number of descriptors associated with the pattern Components, x; - The components of pattern vectors are denoted as x; - Each component x; represent the ith descriptor - The nature of these components depends on the approach used to describe the physical pattern. Example (1): For an example of pattern vector generation, a noisy object is shown in fig. 5.16 (a). The aim is to generate pattern vectors for different types of noisy shapes. By using different descriptors of this noisy object, pattern vectors can be generated in many different ways as below. * Ifeach object is represented by its signature, 1-D signals in the form shown in fig. 5.16 (b) are obtained. (0) ° Ox no 3n mn Sx 3x Jn 2m 42 4 4 4 4 Fig, 5.16 (a) Noisy object Fig, 5.16 (b) Signature of the noisy object Scanned with CamScanner5.32_lmage Representation and Recongnition © If each signature has to be represented by its sampled amplitude values, the signatures should be sampled at specified intervals of @ and denoted as 6, 0,, ....0,. Then, the sampled amplitudes are: ¥1=1(01), x2 =1(02) . © Other than these, for cach signature, the first ‘n’ statistical moments can be computed and used as components of cach pattern vector. Xn= (On) Here, signatures sampled amplitudes and statistical moments are quantitative descriptors of the given noisy object. Thus, pattern vectors are generated in different ways. Example (2): In some applications, structural descriptors are used to generate pattern vectors. One such example is the ‘Finger print recognition’. - The features of the fingerprint are called ‘minutiae’ - The interrelationships between these features are used as the primitive components or descriptors to generate pattern vectors. - Some of such descriptors of fingerprint are abrupt endings, branching, merging, disconnected segments etc. Thus, in addition to quantitative measures, spatial relationships are used to generate pattern vectors. 5.6.2 Strings - Strings are pattems of objects that are generated based on simple connectivity of primitive components. - Usually, these primitives are related with the boundary shape. - These are also known as string descriptions. 5.6.3 Trees - Tree descriptions are used in many applications as they are more powerful ones. - Many hierarchical ordering schemes lead to tree structures. Example: - Foran example, a tree structure is generated for a satellite image of a heavily built downtown area and residential areas as shown in fig. 5.17. Scanned with CamScannerDigital Image Processing 5.33 Image oo Downtown Residential Buildings Highways aI Housing Shopping Highways aN IN malls High Large Multiple Numerous density structure intersections . Single Few intersections Low Small Wooded density structures areas Fig. 5.17 A tree description In the above tree structure, the structural, relationship used is ‘composed of - The’entire image represents the root of the tree The next level shows that the image is ‘composed of a downtown and residential area, - The residential area is ‘com posed of” housing, shopping malls and highways and so on. 5.7 RECOGNITION BASED ON MATCHING Pattern recognition is the process of assi; igning patterns to their respective classes. Recognition based on matching is an approach bas ed on ‘decision - theoretic methods’. 5.7.1 Decision-Theoretic Methods These methods are based on the use of decision Functions that are also called as « discriminant functions. Let, X= (xs,x2, ....¥»)" represent an n-dimensional pattern vector. 1,3, -.--€y be the pattern classes where, W ~ number of pattern classes. The objective of the process is to find a suitable pattern class @, for each pattern based on some decision function. Scanned with CamScanner5.34 Image Representation and Recongnition © Decision Functions The first step is to find W number of decision functions denoted as di(X), d2(X) ... dw(X). The property of these decision functions is: ~ Ifapattern X belongs to class «,, then G(X)>d(X), JHN We FA (536) An unknown pattern X belongs to the i" pattern class if d((X) gives the largest value when substituting X into all decision functions. © Decision Boundary The decision boundary that separates the classes @, and @, is given by the values of X for which 4,(X) =4,(X) (5.37) > . di (X) =4,(X) - di(X) =0 Now, the pattern classification is d,(X)> 0 + for patterns of class «, dj(X) <0 > for pattems of class , 5.7.2. Matching Techniques Recognition techniques based on matching represent each pattern class by a prototype pattern vector. Then the pattern that is closest to this prototype is found based on some predefined criteria. There are two such predefined criteria defined : (i) Minimum distance classifier (ii) Correlation based method 5.7.3 Minimum Distance Classifier - This is the simplest method of matching. - It computes the Euclidean distance between each unknown pattern and each prototype pattern vector. The pattern with smallest distance belongs to the particular pattern class. Scanned with CamScannerDigital Image Processing 5,35 © Prototype Pattern Vector Let, each pattern class is represented by a prototype that is computed by taking the mean vector of all the patterns in that class, i.e., DX Gaba W ae (5.38) 5 xem Where, W — number of pattern classes Nj number of pattern vectors in class w, « Euclidean Distance The Euclidean distance between a pattern and a pattern class or prototype is calculated by D(X) =(|[X-my|, j=h2..W ee (5.39) This is based on the Euclidean norm: |jal| = (a".a)!? © Minimum Distance Now, the pattern with smallest distance is the best match for the particular prototype. It is found as 1 40) =XIm;—Smym;, j=1,2,..W (5.40) Thus, x is assigned to class «, if d,(X) yields the smallest value, * Decision Boundary The decision boundary between classes @, and @, for a minimum distance classifier can be calculated from eqn. (5.37) as di(X) = di(X) — d\(X) substituting eqn (5.40), 1 4,(X) -[x'm, - 7™ | - [x?m, - mj m4] 1 1 1 x! (m,-m,)-> (m,-m,)' (m,+m,)=0 ---(6.4) This eqn. gives a perpendicular bisector of the line segment joining mj and mj. Scanned with CamScanner5.36 Image Representation and Recongnition Advantage: © The minimum distance classifier gives good performance when the distance between mean values is large when compared to the spread of each class with respect to its mean. 5.7.4 Matching by Correlation In this approach, the pattern showing maximum correlation with a pattern class is decided to be best match for that class. © Correlation Theorem The spatial correlation of a mask (x,y) of size m x n with an image f(x,y) is in the form, ce Y=PYoGt). fxs yt) ee (5.42) aaa This eqn, is evaluated for all values of the displacement variables x and y so that all elements of the mask ‘ «’ visit every pixel of the image ‘P. This spatial correlation is related to the Fourier transform via the correlation theorem, f(xy) * WX) © F*(u,v). W(u,v) w---= (5.43) Where F* - Complex conjugate of F. This eqn. results in a two-dimensional circular correlation. © Correlation Matching The normalized correlation coefficient for correlation matching is given by Ly[o6-o] Ur[tew+s, Y+t)-F(x+s, y+t)] “4 {EX[oso-af. LY [ters y+9-Foees, yey] y ~(5.44) yuy)= Where, o- average value of the mask f(x+s, y-+t) - average value of ‘f in the region coincident with o. Here, the values of y(x,y) is in the normalized range [-1, 1]. Scanned with CamScannerDigital Image Processing 5.37 Now, matching is found as below: = When the normalized @ and the corresponding normalized region fare identical, maximum value of y(x,y) occurs. This indicates maximum correlations and hence the best possible match for a pattern with that pattern class. * The minimum correlation occurs when the normalized functions of o and f have least similarity. ‘As, the mask @ can also be referred to as a template, this correlation is also known as ‘Template Matching’ This template matching procedure explained is illustrated in fig. 5.18. Here, Template (@) Image (f) "y! @&y padding Fig.5.18 Template matching procedure Padding(adding zeros on the borders) of the image is done in order to place the mask on the border pixels of the image. ‘The center of the template or mask @ with size mxn is placéd at an arbitrary location (x,y) ‘Then, the correlation at this point is computed by using eqn. (5.44) Now, the centre of the template is moved to the next location and the procedure is repeated. ‘Thus the complete correlation coefficient (x,y) is obtained by moving the center of the template and placing it on every pixel in f. Scanned with CamScanner$.38_Image Representation and Recongnition ~ Finally, the best match is found by selecting the maximum value of (x, y). = It is also possible that there can be more than one matches between @ and f, if they have the same value of (x,y). Drawback: = As normalizing the image involves spatial scaling or image resampling, normalizing for size and rotation is a more complicated problem. Scanned with CamScanner