I'm designing a supervised network that would require to output an image. I'm wondering what are the best metrics to find similarity between the output and actual target image. So far, my best assumption is to calculate the distance of RGB values, and give less weight to the background. Are there better ways?
Widely adopted method to measure similarity between two images is structural similarity:
However, if you aim on training a network that should produce images of desired properties, I would consider GANs framework. In this case, the similarity between the generated and target image is defined by a special discriminator network. Depending on what is your task, there are many GAN sub-types.