The Use of Context in Pattern Recognition
Home
a. Lower Order Markov Chains
b. Hilbert Space Filling Curves
c. Markov Meshes
d. Dependence Trees
Context in Text Recognition
a. A Quick Bit on Compound Decision Theory
b. Dictionary Look-up Methods

## 1.  Introduction

In our everyday lives we end up using context in order to perceive and recognize patterns.  If one would have to obtain an independent solution without the luxury of looking at how the problem was posed, they would be in for a very difficult task.  If one were to interpret how speech were interpreted they would find it highly dependent on the context in which the the words being spoken.  Take for example the utterance "jeetyet."  This would make very little sense unless it were spoken in a high school cafeteria where one friend was asking another if he or she had eaten yet.Duda and Hart

The idea is to consider some entity Z (such as written characters or images).  Z has certain properties when it is viewed on its own versus when we look at it some other context.  The entity Z, can further have even different properties when seen in one context A, versus another context B.Toussaint  There are many examples of how this is relevant.

Consider the word celebration versus the word  CELEBRATION!!!!!! Cleary one example is much more festive than the other based solely on the context that is introduced.  In fact it may even be the first thing that you noticed on the computer screen.  However, one must realize that the context does not lie only in the data but also lies with the "perceiver" in the form of expectations and can depend heavily on culture.Toussaint

Consider the two lines shown below in Figure 1.  Both lines are of exactly the same size.

The lines on the left are of the same length.  However if one were to look at these lines in a different context, like in the one defined on the right, the horizontal line on the bottom appears to longer than the one on the top.  However a copy of the horizontal lines on the left were the ones that were used to create the additional image.  This is known as the Muller Lyer illusion.Toussaint

Systems that analyze patterns by analyzing the data or input information are called data-driven or bottom-up systems. Those that start from overall expectations and work their way down are called conceptually-driven or top-down systems.  However in order to solve difficult problems efficiently while using contextual information it useful to use both bottom-up and top-down processing simultaneously.Toussaint

Now that we know the approaches that can be used we can discuss what problems can be tackled using context in pattern recognition problems.  These problems can be defined in three general categories:  1) disambiguation, 2) error-correction, and 3) filling in the gaps.  Filling in the gaps can be occurs when data is missing (think of data packets that don't get sent in communications systems or data which is changed or destroyed due to noise).Toussaint

This site will discuss who to use contextual information in order to solve problems presented in image classification and text recognition.

Previous: Home
Next:Context in Image Classification