Use TensorFlow to learn a scale-invariant approximation of an image, i.e defined by half-planes & circles, etc.
Learn f(x,y) = (r,g,b) for all pixels of an image.
This is the "scale free" representation of the image. You can then create a new image from this representation at any resolution, or zoom out to see how the representation extrapolates.
To approximate the image, optimally place lines and/or circles in the plane and choose the color for a given pixel based on what side of each line it is on (and/or whether it is inside or outside each circle). The color choosing step is done with ReLu units on top of the lines/circles.
Input image:
Approximated with 100 lines and 500 color choosing ReLu units:
Approximated with 100 circles and 500 ReLu units:
Approximated with 100 lines and 100 circles and 500 ReLu units:
Same as previous approximation, but higher resolution (-x 5)
Line approximation, zoomed out, high res (-b 3 -x 4).
Circle approximation, zoomed out, high res (-b 3 -x 4).
image_learn.py saves the model after each set of epochs, to a .tf file, which includes the input image and all the training parameters.
You can resume training by loading a model with the -m model.tf option, and override the parameters contained therein with the other command-line options.
Run python image_learn.py to see all options.
NOTE: Hit ESCAPE from the display window to shut down cleanly. Program will exit after finishing the current epoch and saving the model. Do NOT just close the window. This is a VisPy bug where the close callback doesn't get called if the window is closed this way.