Logo Recognition

1

Introduction

CONTENTS

1.1 Motivation ................................................................. 1

1.1.1 Shape recognition ................................................. 3

1.1.2 Proposed method .................................................. 4

1.2 Objectives ................................................................. 6

1.3 Assumptions and input data .............................................. 6

1.4 Book organization ......................................................... 7

1.1 Motivation

Shape recognition, such as identifications of characters, industrial parts, air-

planes and logos, is one of the important application areas in computer vision.

A logo typically consists of an iconic or graphic portion and possibly some as-

sociated text. The iconic region may be as simple as a combination of a few

geometric shapes or as complex as a gray scale or color picture. Several ex-

amples of logos are shown in Figure 1.1.

The shape of a logo can vary from simple to complex. Hence, the study,

analysis and recognition of logo shape are both interesting and challenging.

Besides, logos can act as a valuable means of identifying the origin of a doc-

ument. By recognizing the logo, semantic information about the document is

obtained which may be useful to decide whether or not to analyze the textual

parts. Recognizing logos facilitates the recognition of document class and the

analysis of document class. In addition, logo recognition can be adopted to

retrieve and catalogue products according to their logos in many e-business

applications. Staff members of governmental agencies can easily inspect goods

or other items using smart mobile devices empowered by logo recognition

techniques. It is therefore important to distinguish one logo from another.

However, it is not easy to define or detect the similarity between two shapes.

Compared with any other shape recognition activities, logo recognition is more

difficult because logos are complex patterns, consisting of various shapes and

texts.

1

2 Logo Recognition: Theory and Practice

(a) logo12 (b) logo21

(c) logo47 (d) logo50

FIGURE 1.1: Sample logos.

Introduction 3

1.1.1 Shape recognition

Generally speaking, shape recognition, which can be considered as a model-

based matching, can be formulated as follows: label a test shape as one of

the reference models from a finite collection of model patterns. The shape

matching method should be insensitive to geometric transformations, such as

translation, rotation and scale changes. When the number of test and model

shapes is large, the price for brute force matching between the test and model

shapes is high. To cut down the CPU time, one can use shape indexing to

narrow down the scope of the match.

The shape recognition [42, 48, 68, 203, 208, 219] problem can be ap-

proached within five main frameworks [89]: 1) statistical classification, 2)

syntactic/structural matching, 3) template matching, 4) neural networks, 5)

hybrid matching. A brief description of these approaches is given below.

Statistical approach. The statistical approach uses global shape features

(such as moments [145, 162, 209], morphological curvature scale spaces [93]

and Fourier descriptors [144, 207]) to describe shapes, and employs discrimi-

nant functions for recognition. Global features used in the statistical approach

are easy to compute, but the main disadvantage of the statistical approach

is the assumption that almost all of the images must be visible in order to

measure these features accurately.

Syntactic/structural approach. The syntactic/structural approach

uses local structural features like arcs and segments as primitives to repre-

sent shapes. In the syntactic approach, formal grammars are used for shape

representation. The productions of a grammar describe how complex shapes

can be built up from simpler constituents. The recognition process is based on

the concept of formal language parsing. Many different types of grammars and

parsing algorithms for syntactic shape recognition have been proposed in the

past, such as string grammars [21]. As for the structural matching, the basic

idea is to directly represent the models as well as the unknown test image

by means of a suitable data structure (e.g., strings [14, 100], trees [206] and

graphs [40, 182]) and to compare these structures in order to find the simi-

larity between the models and the unknown test image. It requires a formal

measure of similarity between two structural representations. From a theoret-

ical point of view, structural matching can be considered as a special case of

a syntactic approach [22]. Since the syntactic/structural approach recognizes

the shape by means of primitives, it can handle partial occlusion of images.

However, this approach is complex and requires more computational time.

Template matching. Template matching is one of the simplest ap-

proaches to shape recognition. In template matching, a template to be recog-

nized is available. The test image is matched against the stored template using

a suitable similarity measure, such as the Euclidean distance. While template

matching is effective in some application domains [152, 200] and provides a

high recognition rate [18], it has some disadvantages. Since the template is

rigid, it cannot tolerate deviations from the model template. On the other