Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/mnj-tothetop/english-handwritten-characters-recognizer

A handwritten english character recognizer [0-9, A-Z, a-z] made by using a Dataset of 3409 images. Tensorflow, Keras, Scikit-learn, and OpenCV was used to implement the Convolution Neural Network (CNN). Matplotlib and Seaborn were used to visualize the data.
https://github.com/mnj-tothetop/english-handwritten-characters-recognizer

artificial-intelligence convolutional-neural-networks keras matplotlib opencv-python scikit-learn seaborn tensorflow

Last synced: about 2 months ago
JSON representation

A handwritten english character recognizer [0-9, A-Z, a-z] made by using a Dataset of 3409 images. Tensorflow, Keras, Scikit-learn, and OpenCV was used to implement the Convolution Neural Network (CNN). Matplotlib and Seaborn were used to visualize the data.

Awesome Lists containing this project

README 

        
# English-Handwritten-Characters-Recognizer



## STEP - 1: Loading the Dataset



First step I took was to download the dataset.



The labels and the images names were stored in a CSV (Comma Separated Value) file. I used the `read_csv()` fuction of the Pandas library to achieve this.



![image](https://github.com/Mnj-ToTheTop/English-Handwritten-Characters-Recognizer/assets/153396359/ba7c5dd8-43c2-4bc5-bc30-f00c4c42d127)



Looking at the data it can be noted that the first column contains the Image path while the 2nd contains the True Label of the image. There are a total 3409 images.



I then created a Python list `labels` and `filenames` to store the image path of each image along with its corresponding label.



## STEP 2: Preprocessing

To make sure all the images are of the same type and that all the images have the proper input size for our model, I used the `imread()` function of OpenCV library to read the image and the `resize()` function of the same to resize the image to the correct size.



Then I converted both the labels list and filenames list to Numpy arrays. I also used the `LabelEncoder` from the scikit-learn library to encode the the labels to integers (as some of the labels are characters)



## STEP 3: Separating the Train set and Test set

To split the dataset into train and test set, I have used the `train_test_split()` of the scikit-learn library. Once the the train set and test set were obtained, the labels were One-Hot-Encoded using the `to_categorical()` function of Keras.



## STEP 4: Model Building

The model contains 3 Convolution Layers and Classification Layer.



### *LAYER - 1*

This layer contains two `Conv2D()` layers one with 64 filters and the other with 32 filters, with the kernel size = 3 (i.e. 3x3). Following this, a Max Poooling layer is in place, which gives the output shape of `32x32x32`.



### *LAYER - 2*

This layer contains two `Conv2D()` layer each with 64 Filters and a kernel size of 3 (i.e. 3x3). This is followed by Max Pooling with a size of 2x2 which results in an output size of `8x8x64`.



### *LAYER - 3*

This layer is identical to Layer - 2 and results in an output size of `8x8x64`.



`NOTE: I have used padding to avoid shrinking of the data.`


`NOTE 2: The activation function used is Rectified Linear (ReLu)`



### CLASSIFICATION LAYERS:

#### *FLATTEN*

This layer is used to flatten the matrix to a vector. On flattening the matrix the total elements we got: `8 * 8 * 64 = 4096`

#### *DROPOUT LAYER - 1*

To reduce overfitting, we are dropping 0.25 of the neurons.

#### *DENSE LAYER - 1*

A fully connected layer of nuerons with output of shape of `512`

#### *DROPOUT LAYER - 2*

Dropping 0.50 of the nuerons

#### *DENSE LAYER - 2*

Since we have 62 classes to classify, we have kept the the number of nuerons to `62`



## STEP 5: Training and Testing the model

The model was trained using the `model.fit()` function. 30 Epoches were used.

On evaluating the model using the test set, and the function `model.evaluate()` we can see the model has an accuracy of 75.806%



## IMPROVEMENTS THAT CAN BE MADE :

1. A larger data set can be used to train the model more.

2. A larger number of epoch can be used.

3. Changes in the model can also be done to improve the model.



Dataset Taken from:



  @InProceedings {deCampos09, 


  author    = "de Campos, T.~E. and Babu, B.~R. and Varma, M.", 


  title     = "Character recognition in natural images", 


  booktitle = "Proceedings of the International Conference on Computer 


  Vision Theory and Applications, Lisbon, Portugal", 


  month     = "February", 


  year      = "2009", 


}



Kaggel Link: https://www.kaggle.com/datasets/dhruvildave/english-handwritten-characters-dataset