M.S. AAI Capstone Chronicles 2024

A.S.LINGUIST

4

variables are distinct enough to contribute independently to the predictive model, reducing

multicollinearity concerns in model training.

Figure 1

Examples of images in the training set of the ASL alphabet dataset for all the 29 possible classes

Note. The “del” label in the second panel of the last row stands for “delete”.

In order to address potential overfitting and improve model generalization, feature scaling

and data augmentation were applied before model training. Training ASL images were rescaled

(by dividing pixel values by 255) and converted to grayscale, and data augmentation techniques,

such as rotation, horizontal and vertical shifting, shearing, zooming and horizontal flipping, were

employed.

184