aravindbattaje / mlnd_capstone

MLND Capstone project

The capstone project is aiming to perform multi-digit number recognition on street view imagery.

Dependencies

• Python 2.7
• NumPy/SciPy
• Matplotlib
• OpenCV
• Tensorflow

How to run

The project has realized a neural network model to localize and detect digits in any given arbitrary image. Before running, download and uncompress the svhn_multi_digit and svhn_region models in the saved_models directory. Then, populate a directory with images to be evaluated by the model. Here, some sample images are stored in sample/google_street_view_images directory. Then run the following command:

python multi_digit_recognition.py sample/google_street_view_images

For each input image in the directory, the predicted localized region and digit sequence is displayed.

To evaluate the performance of the multi-digit detection model alone, populate already localized images into a directory, here sample/google_street_view_images_cropped, and run the following command:

python multi_digit_recognition_wo_localization.py sample/google_street_view_images_cropped

In the course of realizing the final model, several individual models were trained and evaluated. To run training on, for example MNIST multi-digit model, run:

python mnist_multi_digit_train.py <RUN_NAME>

where <RUN_NAME> can help identify its training logs and keep its model checkpoints in a clean directory. Make sure corresponding datasets have been downloaded and extracted in the manner specified in dataset directory. The SVHN models additionally take -v argument that optionally enables evaluating of validation set during training.

To evaluate a model, make sure the corresponding model checkpoints are present in saved_models directory and then run for example:

python mnist_multi_digit_eval.py

Alternatively, replace the model checkpoints trained with <RUN_NAME> during training in the eval scripts.

Samples

Some correct predictions by the final model:

Some incorrect predictions by the final model: