Psychology 9221B Intro to Neural Networks: Final Project

Architectures for Denoising and deblurring Neural Networks

• Architecture 1 (DN-Net denosing network, DB-Net deblurring network)

  • input layer: 128 x 128 x 3
  • 2 hidden layers:
    • 8 feature maps
    • kernal size 5 x 5
    • stride = 1
    • sigmoid activation functions
  • output layer: 128 x 128 x 3
  • no pooling residual connections or other

• Architecture 2 (deep DN-Net denosing network, deep DB-Net deblurring network)

  • input layer: 128 x 128 x 3
  • 4 hidden layers:
    • 24 feature maps
    • kernal size 5 x 5
    • stride = 1
    • sigmoid activation functions
  • output layer: 128 x 128 x 3

• training details:

  • loss function: mean squared error
  • 100 epochs and stopped optimization if no improvement in validation set after 5 consecutive evaluations