hannesstark / bachelorthesis Goto Github PK

I need much more information about the architecture. I think it can be meaningful, f we have always a short table which contains the successively applied layer (convolutional (and strides), (un)pooling layers, ReLU functions, skip connections as well as input size of images and resulting feature maps). Can you create a document for the latest architecture, which describe this sequence. With a current description we may discuss with matthias a little bit better.

• If data with very wide and different appearances are utilized, the network convergence will be difficult and require larger amounts of training data
• There are millions of parameteres in a deep CNN which require a lot of training data in order to prevent the over-fitting
• Synthetic data augmentation:
  • Multi-resolution
  • Jittering
  • Rotation (images are rotated by $angle \in [-15,15]$ degree)
  • Translation(images are cropped randomly - RGB and DSM)
  • Flip (images are flipped horizontally and vertically)
  • Color space (each band is multiplied by a random value $value \in [0.5, 1.5]$

I have to use such techniques for further work with other students, so i have to implement such data augmentation methods

• Progress?

Initially the network has to work on the images which are totally different to that of the MNIST dataset. Firstly, we have to decrease the resolution of the images by a factor of 4 and using interpolation methods (bilinear interpolation?), to use the proposed architecture.

If the resolution has to be approximately the same like the input, we can decrease the image by a factor of 1.5 or 2 and divide the images in order to get subpatches. This can be a better approach?

Plotting of neural networks for the use in a thesis:

• Latex code for drawing neural networks

Do you want to use such functions to visualize your proposed network?