# Alphabet Classifier

Image classifier for three alphabets(A,T,O), using multilayer perceptron model.

## Getting Started

## Prerequisites

OpenCV is required to generate inpute vectors for training and testing. *How to install-*

pip install opencv-python

#### The Dataset
The dataset contains (15x100) binary matrix, made by flattening (10x10) alphabet images. There is pre-generated dataset, **"dataset.npy"** and along with it **"targets.npy"** for 15 images that are inside "/inputs".

#### Generating the dataset
Generating the dataset for any three alphabets:

#create_dataset.py
images, targets = generate(folderpath)

images is (numofimages x 100) matrix
targets is (number of images x 3 x 1) matrix for 3 alphabets

### Training the network
trainmodel1.py and trainmodel2.py are used to train 1 and 2 hidden layer perceptron model.

*trainmodel1.py* - Contains (100 x 50 x 3) nodes, 100 being the input layer of feature vectors, hidden layer containing 50 nodes, and 3 output nodes generating 1 for every particular image, for my example-
A=\[1,0,0]
O=\[0,0,1]
T=\[0,1,0]
It generates weight matrices (100 x 50), (50 x3) and two bias vectors (1 x 50), (1 x 3).

*trainmodel1.py* - Contains (100 x 50 x 16 x 3) nodes, 100 being the input layer of feature vectors, hidden layer containing 50 nodes and 16 nodes, and 3 output nodes generating 1 for every particular image.
It generates weight matrices (100 x 50), (50 x16), (16 x 3) and three bias vectors (1 x 50), (1 x 16), (1 x 3).

**Activation function used for every node is sigmoid.
Learning rate = 0.001
Epochs=trained with learningrate=0.001
Error calculated using Mean Squared Error**

The above outputs are stored in "inputs#", where # is model number.

## Running the tests
Use main.py

### Getting weights

Loading weights and target:

weights, bias= get_model(model_number)

### Getting testset

Randomly an alphabet image is selected from "/folderpath" and feature vector(1 x 100) is generated for it.

feature_vec, image_name=get_testset("folderpath")

### Running the model
output is (1 x 3) vector:

output=run_model(modelnumber, weights, bias, input)

The output generated and corresponding image_name is printed as output.

A=[1,0,0]
O=[0,0,1]
T=[0,1,0]
These are the ideal values for images of alphabets A, T, O.

### Example outputs
##### MODEL1

|||
| -------- | -------- |
|||
|||
|||
|||
|||

#### MODEL2

|||
| -------- | -------- |
|||
|||
|||
|||

## Built With

* [OpenCV](https://opencv.org/) - Image operations.
* [Maven](https://numpy.org/) - Matrix operations.

## Authors

* **Abhishek Kumar Yadav** - *Initial work* - [abhk943](https://github.com/abhk943)

## Acknowledgments

* [NNandDeepLearning](http://neuralnetworksanddeeplearning.com/chap2.html) - Contains a simple article on error back propagation.