Author: Vincent Yeo

Image Recognition with MNIST

Background:

Handwritten documents are commonly used especially in forms handled by organizations like government agencies or banks etc with individuals. Despite increasing usage of digital platforms that handles online form submission, handwritten form are still a popular means of submission and more can be done to automate the process of form handling and digitalize such documents. One way is by enabling machines to identify chracters and numbers from such documents. This is of course challenging given the unique handwritting of each individuals. Thus the use of neural networks is necessary.

Dataset

The MNIST database contains 70000 images of handwritten digits from 0 to 9, a subset of the NIST database, that has been preprocessed in normalising size and centering of the image.

The authors have offer it to the world for machine learning and pattern recogrnition methods

From LeCun, Y., Cortes, C. & Burges, C. J. C (n.d.). THE MNIST DATABASE of handwritten digits. Retrieved from http://yann.lecun.com/exdb/mnist/

Task:

• Classify the handwritten digits correctly.

Results and Analysis:

The CNN has achieved a 96.2% test accuracy.

Based on each precision score, the model's prediction would be realised truely for

• the digit 0, 97% of the time.
• the digit 1, 98% of the time.
• the digit 2, 97% of the time.
• the digit 3, 95% of the time.
• the digit 4, 96% of the time.
• the digit 5, 95% of the time.
• the digit 6, 96% of the time.
• the digit 7, 95% of the time.
• the digit 8, 96% of the time.
• the digit 9, 96% of the time.

Based on each recall score, the model's prediction for

• the digit 0 has about 99% being correct.
• the digit 1 has about 99% being correct.
• the digit 2 has about 95% being correct.
• the digit 3 has about 96% being correct.
• the digit 4 has about 96% being correct.
• the digit 5 has about 96% being correct.
• the digit 6 has about 97% being correct.
• the digit 7 has about 95% being correct.
• the digit 8 has about 95% being correct.
• the digit 9 has about 94% being correct.

Based on each f1 score, the weighted average of the precision and recall for

• the digit 0 is 98%
• the digit 1 is 98%
• the digit 2 is 96%
• the digit 3 is 95%
• the digit 4 is 96%
• the digit 5 is 95%
• the digit 6 is 97%
• the digit 7 is 95%
• the digit 8 is 95%
• the digit 9 is 95%

Future Works:

• A form would have more than just digit, so more can be done in building a CNN that can recognize characters too
• Position of the character and digits is also important in identifying what fields is that entry associated to. The model would need to identify and sieve out the form fields associated to the handwritten values.
• This would increase the complexity of the model and perhaps separate specialised models can be built to handle the task better.