This repository contains the code necessary to train and test convolutional neural networks (CNNs) for predicting the rotation angle of an image to correct its orientation. There are scripts to train two models, one on MNIST and another one on the Google Street View dataset. Since the data for this application is generated on-the-fly, you can also train using your own images in a similar way. A detailed explanation of the code and motivation for this project can be found in my blog.
You can install all the required packages using pip:
pip install -r requirements.txt
The recommended way to use Keras is with the TensorFlow backend. If you want to use it with the Theano backend you will need to make some minor modifications to the code to make it work.
python train/train_mnist.pyto train on MNIST or
python train/train_street_view.pyto train on the Google Street View dataset. Note that the first time you run the scripts will take longer since the datasets will be automatically downloaded. Also, you will need a decent GPU to train the ResNet50 model that is used in
train_street_view.py, otherwise it will take quite long to finish.
If you only want to test the models, you can download pre-trained versions here.
Note also that the regression models (
train_street_view_regression.py) don't provide a good accuracy and are only included for illustration purposes.
You can evaluate the models and display examples using the provided Jupyter notebooks. Simply run
jupyter notebookfrom the root directory and navigate to
Finally, you can use the
correct_rotation.pyscript to correct the orientation of your own images. You can run it as follows:
You can also specify the following command line arguments: -
-o, --outputto specify the output image or directory. -
-b, --batch_sizeto specify the batch size used to run the model. -
-c, --cropto crop out the black borders after rotating the images.