 Preview: Object Recognition

Patterns and Pattern Classes in Object Recognition

Image processing main steps: • The automatic recognition of objects or patterns is one of very important image analysis tasks.
An object recognition system is like a computer program that tries to identify things in pictures
of the real world.

TYPES OF DIGITAL IMAGE • It does this by using models of objects that it already knows about. This is a tough job because
it's not easy to teach computers to do this. Humans, on the other hand, can recognize objects
PROCESSING ? in pictures without any effort and really quickly.
(1) Low level processing: Primitive operations such as noise reduction, image
Object Recognition
sharpening, enhancement etc. Input and output are images. Central idea: Learning from sample patterns
Methods
(2) Mid level processing: Image segmentation, classification of individual pattern arrangements
objects etc. Here input are images but output are attributes of images for Structural methods
Decision-theoretic methods
e.g. edges of image.
deal with patterns best described by qualitative descriptors
(3) High level processing: It involves making sense of recognized objects and deals with patterns described using quantitative (symbolic information), such as the relational descriptors for
performing functions associated with visions. For e.g. Automatic character descriptors and have patterns arranged in numeric repetitive structures, e.g. staircase). Here the Patterns are
vectors, such as length, area, and texture. arranged in strings or trees.
recognition, military recognition, autonomous navigation etc.

 Patterns and Pattern Classes:  Numeric Pattern vectors (Decision-theoretic methods):

• A pattern is an arrangement of descriptors (or features). • Fisher 1936, performed Recognition of three types of Iris flowers(Iris serosa,
• A pattern class is a family of patterns that share some common properties. Pattern classes are versicolor and virginica by the lengths[x1] and widths[x2] of their petals.
denoted w1, w2, . . . , wN where N is the number of classes.
• Here 3 pattern classes are w1, w2 and w3 corresponding to Iris setosa, versicolor
• Pattern recognition by machine involves techniques for assigning patterns to their respective
classes automatically and with as little human intervention as possible. and virginica. Also there are variations between and within classes.
• The object or pattern recognition task consists of two steps: • Class separability depends strongly on the choice of descriptors.
 feature selection (extraction)
 matching (classification)
There are three common pattern arrangements used in practice:
• Numeric pattern vectors (for quantitative descriptions)
𝑥1
𝑥= 𝑥2[ ⋮
]
𝑥𝑛
• Strings and trees (for qualitative(structural) descriptions)
x = abababa….

 Numeric Pattern vectors (Decision-theoretic methods):  String descriptors (Structural methods):

• Instead of using “raw” sampled signatures, a more common approach is to compute some • In some applications, patterns are best described by structural relationships. eg: fingerprint 1
function, xi=g((r(θi)) of the signature samples and use them to form vectors. recognition.
• These vectors become points in n-dimensional Euclidean space, and pattern classes can be • Primitive components that describe fingerprint ridge properties are:
imagined to be ‘clouds’ in n dimensions.  Interrelationships of print features (minutiae).
• Instead of using signatures, other shape characteristics can be employed (e.g. statistical  Abrupt endings, branching, merging, disconnected segments,…
moments).  Relative sizes and locations of print features
• The choice of descriptors has a profound role in the recognition performance. • For example: Staircase pattern described by a head-to-tail structural relationship. The rule
allows only alternating pattern and It excludes other types of structures but other rules may be
defined.
 Tree descriptors (Structuralmethods):
• A more powerful approach is the use of tree  Decision-theoretic methods:
descriptors. Basically most hierarchical ordering • The basic concept in decision-theoretic methods is the idea of pattern matching
schemes lead to tree structures. based on measures of distance between pattern vectors. Where it includes decision
• In the satellite image example, the structural (discriminant) functions and decision boundaries.
relationship is defined as: “composed of
downtown area and surrounding residential area.
• Let x=[x1, x2,…,xn]T represent a pattern vector.
• For W pattern classes ω1, ω2,…, ωW, the basic problem is to find W decision functions
d1(x), d2(x),…, d W (x) with the property that if x belongs to class ωi:
di(x) >dj(x) for j=1,2,..., W; j# i
• In other words, an unknown pattern x is said to belong to the ith pattern class if upon
substitution of x into all decision functions, di(x) yields the largest numerical value. We
want to classify x, which is a pattern. We are given a finite set of classes of objects. We want
to categorize the pattern x into one of the classes. To do so, we apply x to all decision
functions, and categorize x to the class of best fit.

 Decision-theoretic methods (cont.):

•The decision boundary separating class ωi from class ωj is given by the
values of x for which di (x) = dj (x)or  Types of Decision-theoretic methods:
• Matching: an unknown pattern is assigned to the class to which it is
• If x belongs to class ωi: closest with respect to a metric.
–Minimum distance classifier (Computes the Euclidean distance
between the unknown pattern and each of the prototype vectors).
– Correlation (It can be directly formulated in terms of images)
• If x belongs to class ωj:
• Optimum statistical classifiers (A probabilistic approach to recognition)

•Neural networks (nonlinear computing elements (called neurons)

organized as networks)

 Matching and Minimum distance classifier:  Minimum distance classifier(cont.):

• Matching is an important concept in computer vision and object recognition. Images of the • The decision boundary between classes ωi 2
same item can be taken with any angle and any lighting condition and any scale. The simplest and ωj is given by:
approach of matching is the minimum distance classifier. Matching represent each class by a
prototype pattern vector. The prototype of each pattern class is the mean vector:

• The surface given by above equation is the

• Using the Euclidean distance as a measure of closeness: perpendicular bisector of the line segment
joining mi and mj.
• For n=2, the perpendicular bisector is a line, for
• We assign x to class ωj if Dj(x) is the smallest distance. That is, the smallest distance implies the n=3 it is a plane and for n>3 it is called a
best match in this formulation. hyperplane. The two classes, Iris versicolor and Substitution of any pattern vector from class ω1 would
• It is easy to show that selecting the smallest distance is equivalent to evaluating the functions: Iris setosa, denoted ω1 and ω2, respectively, yield d12(x) > 0. Conversely, any pattern from class ω2
have sample mean vectors m1 = (4.3, 1.3)T and would yield d12(x) < 0. In other words, given an
unknown pattern belonging to one of these two classes,
m2 = (1.5, 0.3)T. The decision functions are
sign of d12(x) to one of these two classes, sign of d12(x)
and assigning x to class ωj if dj(x) yields the largest numerical value. This formulation agrees with would be sufficient to determine pattern's class
the concept of a decision function. membership.
 Minimum distance classifier (cont.):
• In practice, the classifier works well when the  Structural Methods :
distance between means is large compared to the
spread of each class. This occurs seldom unless the The techniques discussed in decision theoretic methods deal with patterns quantitatively and largely ignore any structural
relationships inherent in a pattern's shape. That is Decision theoretic methods focus on numbers and don't pay much
system designer controls the nature of the input. attention to the waythings are arranged in a pattern.
• An example is the recognition of characters on bank
checks. - American Banker’s Association E-13B font • On the other hand, structural methods are all about recognizing patterns by paying close attention to how things are
arranged (qualitatively) or connected in the pattern. The structural methods discussed here, however, seek to achieve
character set. pattern recognition by capitalizing precisely on these types of relationships.
• Characters are purposely designed on a 9x7 grid to
Types of Structural Methods :
facilitate their reading. Also the characters are • MatchingShape Numbers:
stylized to maximize the difference between them. Matching Shape Numbers is a structural method in object recognition that involves comparing objects based
• The font was designed to be readable by humans, on their first difference chain code representations to determine if they match.
but the overriding purpose was that the characters
should be readable by machine, quickly and with • String Matching
high accuracy. String Matching is a structural method in object recognition that involves comparing the object's boundary
encoding as a string with a reference template to determine if they match.
• The waveforms (signatures) are different for each
character.

 Matching Shape Numbers : • LetMatching Shape Numbers (cont.) :

a and b denote two closed shapes which are represented by 4-directional chain codes and
•The degree of similarity, k, between
s(a) and s(b) their shape numbers.
two shapes is defined as the largest
• The shapes have a degree of similarity, k, if:
order for which their shape
numbers still coincide.
− Reminder: The shape number of
a boundary is the first difference of
• This means that the first k digits should be equal.
smallest magnitude of its chain code
• The subscript indicates the order. For 4-directional chain codes, the minimum order for a closed
(invariance to rotation).
boundary is 4.
− The order n of a shape number is
• Alternatively, the distance between two shapes a and b is defined as the inverse of their degree of
defined as the number of digits in its
similarity:
representation.
• Examples. All closed shapes of
order n=4, 6 and 8.
•First differences are computed by
• It satisfies the properties:
treating the chain as a circular
sequence.

 Matching Shape Numbers :  String matching :

3
• Suppose that we have a shape f and want to
• Suppose that two region boundaries, a and b,
find its closest match in a set of five other are coded into strings denoted by a1a2a3 …an
shapes (a, b, c, d and e), as shown in Fig.(a). and b1b2b3 …bm.
• Suppose that the shapes are identical upto •Let p represent the number of matches
degree 8, with the exception of shape a
whose degree of similarity with respect to all between the two strings.
other shapes is 6. Shape d has degree of − A match at the kth position occurs if ak = bk.
similarity 8 with respect to all others, and so on. • The number of symbols that do not match is:
• Shapes f and c match uniquely, having a
higher degree of similarity than any other
two shapes.
• If ahad been an unknown shape, all we could • A simple measure of similarity is:
have said using this method is that a was
similar to the other five shapes with degree of
similarity 6.