MICRST Competition

This is the repository accompanying the article: From MNIST classification to Intelligent Check Processing

Quickstart

Use the introductory notebook

Step by step:

1. Prepare the dataset config:

   There is a basic example in ..blob/master/dataset/generator.py#L12
   You can change images dimensions and probabilities of randomly applied image transformations, e.g. setting:
   'blur_probability': 0.5
   will generate a dataset where 50% of samples have blurring applied.

   To preview the dataset use the

   dataset.generator.make_preview(config)
   

   

2. Generate the full dataset:

   Declare total number of sumples and execute the function:

   number_of_samples = 60_000
   dataset.generator.make_dataset(
       config=dataset_config,
       number_of_samples=number_of_samples
   )

   This can take anything from 20 - 90 minutes, depending on the machine you're using and number of samples you choose and will create a new directory in the ./tmp folder of this repository (can change in the config) containing all all of the images and one labels.csv file with information about the ground truth character for each image.

3. Load dataset with PyTorch loaders:

   Using the same config object you can make training and validation data loaders:

   train_loader, test_loader = dataset.dataset.make_dataset_loaders(
       config=dataset_config,
       batch_size=batch_size
   )

4. Train models!

   We are including a neural network model with 3 convolutional layers, 2 fully connected and batch normalization. Is pseudo-python the training procedure boils down to:

   model = models.DeepMicrClassifier()
   for batch in train_loader:
       # Load the data
       image_data = batch['image']
       target_labels = batch['label']
   
       # Feed it to the model
       output = model(image_data)
   
       # Calculate error
       loss = your_loss_function(output, target_labels)
   
       # Back propagation
       loss.backward()
       
       # Gradient descent
       optimizer.step()

   Running the notebook as is should give you 90% accuracy on the validation set.

   

5. Experiment!

   This is clearly a toy example - we encourage you try and beat our baseline. Ideas for improvements:

   • make the model deeper
   • add data augmentation
   • remove convolutional layers
   • remove fully connected layers
   • train multiple models and try the ensemble approach

Dataset structure

The tree arrangement of the data directory:

└── MICRST
    ├── image
    │   ├── MICRST_1000000.png
    │   ├── MICRST_1000001.png
    │   ├── .
    │   ├── ..
    │   ├── ...
    └── labels.csv

And the structure of labels.csv:

label,path
7,./tmp/MICRST/image/MICRST_1000000.png
D,./tmp/MICRST/image/MICRST_1000001.png
.
..
...

Where the label column is the MICR character in string format.