Sign-Language-MNIST

Final project for FOUNDATION OF MACHINE LEARNING exam (Master's degree)

Create a robust classifier to recognize American Sign Language hand poses.

Data

28x28 pixels images of hand poses:

• 24 categories: full English alphabet excluding J and Z which require motion.
• 27455 training images:
  • 80% actual training.
  • 20% validation.
• 7172 test images.

Base architectures trained

• ”LeNet5” (1 architecture): LeNet5 architecture from original paper.
• ”Classifier 2” (12 architectures): CNN with 2 convolutional layers.
• ”Classifier 3” (24 architectures): CNN with 3 convolutional layers.

”Classifier 2” and ”Classifier 3” architectures generated by varying:

• dropout layer positions;
• number of neurons in hidden layer.

TOTAL: 37 architectures