Home > Mobile >  Is it possible to train except for a specific area when implementing cycleGAN?
Is it possible to train except for a specific area when implementing cycleGAN?

Time:02-25

Recently, I am conducting research to create fake images using cycleGAN. The figure below is an example of the cycleGAN result. It can be seen that all areas are changed as in the image. However, I want to change the image only for a specific area. Is it possible to change only the image of the part except for the red marked part (disease part) using cycleGAN?

enter image description here

CodePudding user response:

I think what you describe is possible. You would need to alter the cycleGan pipeline. To focus the representational capacity of the network on everything other than the red marked part you could replace the red marked part with the same part from the original image, this would enforce a focus on the other regions. When taking the cycle consistency pipeline shown below, I'm referring to adding the original image parts after e.g. Generator A2B.

Cycle consistency loss pipeline

  • Related