As described in section 2, In this experiment, we randomly sampled 100 pairs of images from the BSD100 dataset [26], and sum each pair. For each image pair we trained a DIP to learn the summed image and each of the individual images:
The mean MSE loss between the output of the network and the image, during the first steps of the training phase is:
Note the larger loss and longer convergence time of the mixed image, compared to the loss of its individual components. In fact, the loss of the mixed image is larger than the sum of the two individual losses.
We performed a similar experiment for non-overlapping image segments. We randomly sampled 100 pairs of images from the BSD100 dataset. For each pair we generated a new image, whose left side is the left side of one image, and whose right side is the right side of the second image. We trained a DIP to learn the mixed image and each of the individual components.
The mean MSE loss is:
