Machine Learning – Spring 2009 – Project 1
Due Date: March 6, 2009
Consider the images of handwritten characters in Figure 1.
Figure 1. Some handwritten character images.
Without looking at the figure on the next page, how well can you identify the characters? (Your answer
should be “Not very well”). However, if you look at Figure 4, the identities of the characters should be
obvious. This observation leads to a somewhat paradoxical situation if we want to use machine learning
to write a program that reads handwritten words, i.e. we can’t read the characters without reading the
words and we can’t read the words without reading the characters! One response to this is have the
program segment an image of a handwritten word into many small segments, each hopefully consisting of
a single character or something less than a character. Then given a string representation of a handwritten
word, use dynamic programming to find the best way to put the segments together to match the string and
assign the match a score. The match score is built from scores assigned to individual segments as
depicted in Figure 2.
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
Input
Image
Primitives
Best
Match
to
"Richmond"
Best
Match
to
"Edmund"
Image Segmentation
R=53
i=27
c=52
h=61
m=70
o=43
n=61
d=88
E=12
d=79
m-85
u=25
n=61
d=88
Figure 2. Dynamic Programming Approach to Handwritten Word Recognition
This changes the character level problem from asking the question “Which character does the input image
represent?” to “How much does the input image look like it represent a specific character (such as u or
v)?”. This approach actually works quite well and is the basis of several (at least two that I know of
