Handwriting Recongnition Using
Machine Learning

Our Project

Cinque Terre

Model

Our model consists of five layers i.e. two convolution-pooling layers with 20 and 40 feature maps respectively and two-by-two max-pooling and one fully-connected hidden layer with 100 neurons.

Cinque Terre

Image Processing

Processed handwritten images to resemble the images of training-set.

Cinque Terre

AlphaNumeric Recognition

Our model takes images of handwritten text as input and predicts the output.

Cinque Terre

Segmentation

If input image has many lines, we have developed a segmentation algorithm which clips out every character and then forwards it to model.

Details Of The Model

  • Convert the input image into 28x28 matrix
  • After input layer there are two convolution layers with 20 and 40 feature maps and a receptive field of 5x5 each
  • In both the convolution layers Max Pooling is done using a window of 2x2.
  • Then there are two fully connected layers with 2560 and 100 neurons . Tanh() activation is used for both the layers.
  • At last comes the output layer with 36 neurons as we can have output in 36 classes.

Our Project

  • About The Project

    We have worked on a program that can recognize handwritten text which contains english alphabets and digits . Just input the image and you will the get the text written on it.

  • Abstract

    We have used neural network which learns the parameters so that it can learn the scriblings of the alphanumerics. The learning part is done using a character dataset. When we give a input , the program crops out each character in the order they are written and fed to the network for recognition. And there we have the text.

  • Model

    We have worked on a neural network which is trained by a character dataset to recognize handwritten digits and alphabets within accuracy of 92%.

  • Image Processing

    Worked on image processing by using scaling, normalization and translating the image so that it resembles MNIST trainset images.

  • Improved Model

    Improved our model by using Chars74k dataset and by adding two convolution-pooling layers and one fully-connected hidden layer with 100 neurons which increased accuracy over 90%.

  • Segmentation

    Developed a function for segmenting the image into individual characters, applying above methods to get the image recognized.

  • Work Accomplished

    After segmentation, the characters get recognized including SPACES and then furthur words. After the words are recognized, they are corrected if some letters are wrongly recognized using a dictionary spell-corrector. Given a paragraph of multiple lines, finally it can be recognized.

  • Scope of Improvement

    To make an interface to ease the communication between user and the program. To recognize CAPTCHA.

  • Sources

  • Thank
    You
    :-)

Our Amazing Team

Pratham Kumar Verma

Pritesh Kumbhare

Chaitanya Dhawan

Bishal