Pattern recognition using singular value decomposition

Source code is under src and my images are in run. The document report.pdf is my paper on the topic, produced using LaTeX in the report folder.

Abstract

This paper presents a linear algebra method for pattern recognition. Singular value decomposition is used to distinguish single handwritten digits. I implement a method for giving test images to the program and examine the goodness of recognition when accounting for different number of variations in the training data.

Run in MATLAB command window:

Go to src

>> cd [your-path-to-svd]/SVD-pattern-recognition-master/src/

Run main.m

>> main

Welcome! This is the pattern recignition bot beep boop
------------------------------------------------------
Write "test" for testing one number or "graph" for data collection: graph
Filename of your digit image: image3.png
Integer value for cut-off parameter k: 30

or

Welcome! This is the pattern recignition bot beep boop
------------------------------------------------------
Write "test" for testing one number or "graph" for data collection: test
Filename of your digit image or "r" for a random number: r
Integer value for cut-off parameter k: 20

• The first time the program is run a variable named U should not exist in the workspace (save your work and start off with a clean slate clear; clc).
• The training data U takes approximately 30 seconds to finish, its size is ~30Mb.
• Going forward or rather running the recognition multiple times the U will not be recalculated since the program checks for its existance in the base worksapce.

More on how to run using your own images

The run/ folder contain readily available test .png images and you can create your own in any old graphics editor.

Image requirements

• Square shaped: height px = width px
• White background: rgb(255,255,255)
• Black digit: rgb(0,0,0)
• Placement: In the run/ folder!

Demo output:

>> main
Welcome! This is the pattern recignition bot beep boop
------------------------------------------------------
Write "test" for testing one number or "graph" for data collection: test
Filename of your digit image or "r" for a random number: image7.png
Integer value for cut-off parameter k: 42
Relative residual = 0.44336
Predicted number = 7

This is what the input image looked like:

And this is how my pattern recognition program interprets it so that it can be compared with the NIST data:

Now plotting relative residual vs parameter k:

>> main
Welcome! This is the pattern recignition bot beep boop
------------------------------------------------------
Write "test" for testing one number or "graph" for data collection: graph
Filename of your digit image: image7.png
Integer value for cut-off parameter k: 30