Change the gui to allow the definition of the scale steps to apply. On
each scale, the image will be rescaled by the defined factor and
additional patches will be computed. Works best with using mean
accumulation instead of maximum accumulation.

Add a gui checkbox to select using mean accumulation. With this, the
final crack probability per pixel is the mean of all predicted patches
containing this pixel. If mean is not used, maximum accumulation is
used.

robust crack detection. Should be somehow configured in the GUI

- add one last row and column to always check the border regions
- add option to use maximum CNN response instead of mean for pixel prediction

Mobilenet and Inception v3 retrained on SDNET2018