GENERAL
The main goal of the project work is to implement a widely used
classification method and test it with fairly simple data. The work is
carried out with Matlab software. In this work, the data consist of
handwritten digits.
FILES
As the main purpose of the work is not to learn how to use Matlab but
to get more familiar with pattern recognition methods, some of the
Matlab code is provided. Your task is to code the recognizer.
Data
The handwritten digit data are stored in Matlab files
here.
The files contain digits written by two or four writers. Use
either one of the data sets in your experiments. The digits are stored
as two two matrices: one-stroke digits are in the different matrix than the
digits consisting of two strokes. The size of a digit matrix is N*2*n,
where the first dimension is the number of digits, the second
dimension indicates either x- or y-component of the data point, and
the third dimension is the number of the data points. The data points
are the positions of the pen point on the writing surface. Each stroke is
presented with 50 evenly spaced data points. The labels of the digit
samples are given in separate vectors.
You can have a look on some of the unprocessed digit samples at here.
Matlab functions
You can preprocess and normalize the digits with the following
functions: Decimate, Centralize, and
NormalizeSize. The first function reduces the data by keeping
every n:th data point and abandoning the rest of the data points. The
second function moves the center of the bounding box drawn around the
digit into to the origin of the coordinate system. The third function
scales the character so that the height of the bounding box is set
to a constant value and the aspect ratio remains unchanged. With
function DivideSetsLottery you can divide the digits into two
sets. Function DtwDistance calculates the Dynamic Time Warping (DTW)
distance between two digits. It also returns the difference between the
digits which you can use in the Learning Vector Quantization algorithm.
DTW is a nonlinear curve matching method commonly used in handwriting
and speech recognition. You can find more information on the
DTW-algorithm from the comments of the function's Matlab code.
You can get the functions from here: Decimate.m, Centralize.m, NormalizeSize.m, DivideSetsLottery.m, DtwDistance.m.
EXERCISES
Preprocessing methods
If you reduce the number of data points you can decrease the time
consumed in the evaluation of the DTW-distance significantly. However,
you will lose some valuable information. How many data points can
represent a stroke with a sufficient accuracy?
Normalization methods
The digits have been written without any specific constrains on the writing
style. The variations caused by different personal styles of
writing can be covered, in some extent, by the DTW-algorithm used
for comparison of the digits. However, the size and location variations
of the digits have to be normalized before they can be matched. How do
personal writing styles differ from each other?
Feature selection
The data point sequences are compared as such and there is no need for
a specific feature extraction.
Recognition methods
Implement one of the following recognition methods: 1) the k Nearest
Neighbors (knn) classifier, 2) Learning Vector Quantization (LVQ)
classifier. For both methods, use DTW-algorithm for comparing the
digits. Experiment with different values of k and prototype set
sizes. Use linearly decaying learning coefficient for LVQ.
Do not use same digit samples for learning and testing the recognition system.
Reporting
In the report, give a brief description of the recognition method
applied, justify all the selection you have made, and give the
recognition results of the best recognizer (calculate the total error
rate and occurrences of all error types). Also, include your Matlab
code with comments.
This page is maintained by elia.liitiainen@hut.fi.