Learning Iterative Image Reconstruction - Hierarchical Neural Networks for Image Interpretation - page 178

Input 2 3 5 10 Target
Fig. 9.4. Iterative super-resolution. The activity of the network's output feature array is shown
over time along with the low-resolution input and the high-resolution target. The initial output
is refined as the number of iterations increases.
Fig. 9.5. Differences between output of the recurrent super-resolution network and the target.
Large deviations occur because the choppy edges of the targets are approximated by smooth
contours.
reproduced, but have been replaced with smooth edge segments. For this reason, the
reconstructions frequently have a higher perceptual quality than the targets.
Figure 9.6 shows the contributions to the activity of the output feature cells after
ten iterations. One can see that the feature cells belonging to the lines are excited
via the backward projections, while their neighborhood is weakly inhibited by the
Layer 1 features. The backward influence on the rest of the image is weak. The
influence of the lateral projections is strongly inhibitory everywhere, except at the
lines, where it is excitatory.