ML results

Handwritten Alphanumeric model

Iteration 1:

Model accuracy - 99.80%
Dataset trained on - Existing dataset

Iteration 2:

Model accuracy - 93.90%
Dataset trained on - Existing dataset + NIST misclassifications (after inference using old model)

Iteration 3:

Model accuracy - 94.10%
Dataset trained on - Existing dataset + manually collected dataset (~1K images)

Handwritten Digits model

Iteration 1:

Model accuracy - 99.90%
Dataset trained on - Existing dataset

Iteration 2:

Model accuracy - 99.70%
Dataset trained on - Existing dataset + NIST data + inference on production dataset (~50 images)

Iteration 3:

Model accuracy - 98.30%
Dataset trained on - Existing dataset + NIST missclassifications (from previous model in Iteration 3) + manually collected dataset (~8.6K images)

Sample dataset images

Existing dataset

NIST dataset

Manually collected

Some unhandled misclassifications