Why OCR?

OCR has wide-ranging implications in many industries.

Anywhere there is a need to convert handwritten text into machine-encoded text, OCR can be used to reduce errors and increase speed.

 Choosing a Deep Learning Model

• CNN Based Computer Vision Works Best

CNN Recap

Process

• Get Data - MNIST
• Clean and Pre-process - Done!
• Model - Done!
• Test - Demo
• Improve?

Process

• Model - CNN
• Clean and Pre-process
• Get Data

Learning Attributes Instead of Words – Unigram Approach

Consider the case of a training set of size 1,000. The word “SLEEP” may appear only twice, but attributes such as “does the word contain the unigram ‘S’ in the first half of the word could occur multiple times – which is great for CNN's.

Similar words may confuse the network - Consider the words “KIDS” and “BIDS”. A “KIDS” word image is a negative sample for the “BIDS” category, although a large part of their appearance is shared. This similarity between some categories makes a category based classifier harder to learn, whereas an attributes based classifier uses this to its advantage.

STEP 1: Getting and Cleaning Data

• IAM Dataset

• 657 writers contributed samples of their handwriting

  1'539 pages of scanned text

  5'685 isolated and labeled sentences

  13'353 isolated and labeled text lines

  115'320 isolated and labeled words

 Step 2 -Training and Preprocessing 

• Normalization – Convert to grayscale – Divide each pixel by 255.
• Label Encoding – One hot vector.
• Image augmentation

 Step 3-Training

• Monitor Validation Loss
• Reduce Learning Rate
• A word on accuracy

Step 4-Testing and Improving

• What if the results are not that good?
• Monitor loss function
• Why is accuracy so high
• Should we consider accuracy or some other score?
• Does not generalize well even though the training and validation accuracy is good.
• Overfitting
• After training and testing accuracy is improved, how do we generalize to other areas?
• Pre-processing of form data
• What to do with the line
• How to filter blue values
• Pre-processing issue, still not working/generalizing well outside of IAM database

• Visualizing the neural network

DEMO