home | download | howto | module docs | git

cnngenin: generate training data as input for nn optimisation

this takes a sequence of input and reference image pairs and prepares them for neural network training such as kpn-t over the course of an animation.

every frame will present a different tile with different rotation and exposure, to generate more variation that present in the dataset.

this module is meant to be connected to a chain of input that stems from sequence data (use the %04d pattern in the filename input of i-jpg, say).

connectors

• imgi the input image to be processed
• refi the reference image to be compared to
• imgo a random tile of the input image
• refo the corresponding tile of the reference image
• noiseab the generated noise profile values to pass on to training

parameters

• generate switch randomisation of input on or off, or add synthetic noise