Character Classification

The repository is an attempt to solve the challenge given by MIDAS for their internship program.

All the code is in character_classification.ipynb and some of the other items which were attempted is mentioned in experiment_log.ipynb

Pre-requisits

1. Modules needed to run the notebook are all mentioned in env.yml file. I use Anaconda and you can create the environment I used for this project using

   conda env create -f env.yml

2. The notebook assumes that the data-folders for the task-2, subtask-1 and 3 are present in the location.

3. The network and models are fairly small and thus have been pushed to the repository. Model checkpoints are all present in directory /models

4. The test data for subtask-1 should also be present in the location where the notebook is.

Additional Info

• The static folder contains images which are used in the notebook.

• predict for class MIDAS-CNN is not written atm because it was not need but it will be useful for prediction and should be straight forward to write.

• Data for subtask-1 and for subtask-3 are linked.

• Data for testing subtask-1

Test evaluation

The test accuracy for sub-task 1 for my network is 63.44%

Where the model lacked, involved a lot of possible confusions to naked-eye is well like capital and small alphabets, other confusions include confusion between o, 0, O, Q; confusion between B, 8 etc.

Atleast for small and capital context will come in extremly handy therefore predicting the character knowing a sequence will be something that will help the model in detecting hand-written characters.

(Note: To use an end-to-end approach for character recognition of hand-written text, will work on it soon)