Handwritten character Recognition

Introduction

In this repository I used images of all alphabets and all digits. The images are used to train Convolutional Neural Network. This model was used for prediction. The datset contains 36 classes: all digits and all lower case alphabets. The model gave about 76%.

Preprocessing Data

A filled form image was given with black boxes at top letf corner. Using the box has reference a mathematical formula for derived to extract images. Thresholding and dilation were applied on these images to change the color with threshold and to maximize the image for prediction using CNN.

Network Graph

The CNN model

Layer	Description
Convolutional Layer	Filter Size:2*2 Filters:32
Convolutional Layer	Filter Size:2*2 Filters:32
Max Pooling Layer	Pool Size:2*2
Convolutional Layer	Filter Size:2*2 Filters:64
Convolutional Layer	Filter Size:2*2 Filter:64
MaxPooling Layer	Pool Size:2*2
Flatten	Reshape 4D into 2D
Dense	Neural Network Layer, Neurons:128
Dropout	Avoid overfitting, rate:0.2
Dense	Neural Network Layer, Neurons:36

Training

1. Loss function: Cross Entropy
2. Optimizer: momemtum Gradient descent
3. used stochastic gradient descent

Prediction

The model was saved. The saved model was used for prediction with the images obtained from during preprocessing.