ML Final Project: Aidan Aug, Alan Zhang, Shreya Wadhwa, Trisha Karani
The purpose of this project is to create a CNN Classification for ASL Hand Symbols.
Augmentation.ipynb: contains code for loading raw dataset (handgesturedataset_part1), preprocessing, and creating augmented dataset with blur, scale, rotate 30 degrees, vertical flip, and translation transformations. Creates FinalImages folder with preprocessed and augmented images. Additionally, contains code for applying multiple augmentations to one image, creates Mult_Augments folder.
FiveConvLayerCNN.ipynb: contains code for training and testing AlexNet model on normal images and augmented image dataset, as well as prediction on multiple augmentation dataset.
FiveConvLayerCNN-Features.ipynb: contains code for training and testing AlexNet model with hand-to-back and convexity features added to linear layer on normal images and augmented image dataset, as well as prediction on multiple augmentation dataset.
SimpleCNN.ipynb: contains code for training and testing LeNet model on normal images and augmented image dataset, as well as prediction on multiple augmentation dataset.
SimpleCNN-Features.ipynb: contains code for training and testing LeNet model with hand-to-back and convexity features added to linear layer on normal images and augmented image dataset, as well as prediction on multiple augmentation dataset.
convexity.py: contains code for determining and saving the convexity feature on raw images.
HandToBack.ipynb: contains code for determining hand-to-back ratio feature on raw images.
Final_Project_Submission.ipynb: final jupyter notebook containing code and discussion of findings.
FinalImages: folder of preprocessed normal images and augmented images.
Mult_Augments: folder of images with 0-3 randomly-determined augmentations applied (per image).
features: folder of .npy data of features for images.
handgesturedataset_part1: Raw image data of ASL hand symbols for this project. One of five folders from the Massey University hand symbol dataset.
models: folder of saved trained models.