Disposisjon CNN

• Preprosessering
  • Sørge for riktig format på bildene
  • STN / cropping av bildene for å redusere støy
  • Prøve ut forskjellig Data Augmentation teknikker på bildene og se om det gir bedre klassifisering: https://github.com/joshwadd/Deep-traffic-sign-classification#data-augmentation
• CNN
  • Forslag til modell (Bør prøve ut forskjellige) Denne har Spatial Trandformer Networks som vi ikke har hatt om, men brukes til å øke geometrisk invarians ved feks å croppe bildet. "It can be a useful mechanism because CNNs are not invariant to rotation and scale and more general affine transformations" (Pytorch.org)
    • I alle fall et par lag med
      • Conv
        • The objective of the Convolution Operation is to extract the high-level features such as edges, from the input image. (towardsdatascience.com)
      • Max Pool
        • Max Pooling returns the maximum value from the portion of the image covered by the Kernel (Conv) (towardsdatascience.com)
      • ReLu
        • ReLu layers (Nair & Hinton, 2010) are made up of neurons that apply the activation function f(x)=max(0,x), where x is the input to a neuron. These layers enhance the non-linear properties of the network,including the decision function, without affecting the learnable parameters of the convolutional layer. (Arcos-García, Álvarez-García & Soria-Morillo)
  • Classification — Fully Connected Layer (FC Layer)
    • Softmax: Over a series of epochs, the model is able to distinguish between dominating and certain low-level features in images and classify them using the Softmax Classification technique. (towardsdatascience.com)
  • Optimalisering
    • Kjør tester på følgende og varier modellen, noter accuracy: SDG, RMS-Prop, Adam