This is my approach to American Sign Language Recognition (Alphabet level only) using MNIST ASL Dataset and a small ConvNet, The goal here was to train a model with high accuracy using CPU only on a small dataset. For simplicity, I used Keras with a TensorFlow backend, I also used Numpy for matrix manipulation, Pandas to read the data properly and Matplotlib for plots and visualization.
the two CSV files contain all the image data, each row is basically a flattened image, there is a train CSV for training the model (train + validation) and a test CSV to evaluate how the model does on unseed data.
- Download the 2 CSV files, and the .ipynb file.
- make sure you have the latest TensroFlow version installed, this notebook was created using this docker image :
tensorflow/tensorflow:nightly-devel-py3 - make sure you have Numpy, Pandas, and Matplotlib
- access "CNNonASLofMNIST.md" this file will show you a nice version of the notebook.