KNN Classifier

• Specifications:

1. Train folder that contains folders*, each folder will contain the dataset to train the model.
2. Test folder that contains folders*, each folder will contain the dataset to train the model.

• Preprocess:

Each sub folder in the train/test set goes through the same preprocess steps.

1. converting to gray scale
2. if the width != height, add padding to the small one (white border) to make the picture a square so that width = height
3. resizing to 40X40

The program used the ration of 90:10 for the train/validation sets

• Train:

Giving the model the preprocessed dataset and train it using:

1. Chi Square distance function
2. Euclidean distance function

The program will take the best trained model with K neighbor from 1 to 15 (usually the best is between 7 to 9).

• Test:

After training will test the best model using the preprocessed test dataset.

• Result:

Will save a csv file that contains confusion matrix and classification report. Can be changed to save the best model.

*folders = letter folders (0 = a, 1 = b and so on).
Example of how the file tree should look like:
	-root
		- Train
			- 0
				- pic1.jpg
				- pic2.jpg
				- pic3.jpg
			- 1
			- 2
			...
		- Test
			- 0
			- 1
				- pic1.jpg
				- pic2.jpg
				- pic3.jpg
			- 2
			...
			
		- preprocessed.py
		- knn_classifier.py

Installation

python version 3.7

libraries needed:

• opencv
• numpy
• os
• argparse
• pandas
• scikit-learn
• scikit-image
• datetime
• random

How to use?

1. open terminal/CMD
2. pre process TRAIN and TEST folders -> python preprocess.py [--train TRAIN_PATH] [--test TEST_PATH].

2.1. it will be output to "processed_train" or "processed_test". 3. start the classifier -> python knn_classifier.py .

3.1. output -> csv file that will hold both knn distance results.

*Note: can change the code to save the model and use it in various ways.

Credits

Aviel Cohen - linkedin

Noah Solomon - github, linkedin

Emilia Zorin - github, linkedin

Vitaly Nechayuk - github, linkedin

Dataset contributors

The dataset used in this project was HHD_v0 HHD_v0 was divided into TEST/TRAIN sets which included all 26 sub directories related to each hebrew letter. The dataset, HHD_v0, was developed by I. Rabaev, B. KurarBarakat, A. Churkin and J. El-Sana for their paper HHD_v0