Learning Iterative Image Reconstruction - Hierarchical Neural Networks for Image Interpretation - page 186

20

TRN
TST

TRN
TST

1

15

0.1

10

0.01

5

0 0.001

2 4 6 8 10 12 14 16 18 20 2 4 6 8 10 12 14 16 18 20

(a) (b)

iteration iteration

Fig. 9.16. Filling-in of occlusions - performance over time: (a) mean squared error; (b) mean
square delta output.

can do is to produce a reasonable guess about the portion of the digit behind the
square. No perfect reconstruction should be expected.
The vanishing changes of the output units prove that the network activities in-
deed converge to an attractor even when iterated further than the twelve time steps
it was trained for.

9.4 Noise Removal and Contrast Enhancement

Low image contrast, a varying background level, and noise are common when cap-
turing real-world images. While some of these sources of degradation can be re-
duced by controlling the setup, e.g. by providing homogeneous lighting, other fac-
tors cannot be compensated for. One example is the dark structured paper of larger
envelopes that leads to degraded images of the address or other labeling. Image pro-
cessing can try to reduce noise and improve contrast in order to ease subsequent
recognition steps. The challenging aspect of this task is to separate the noise one
wants to remove from the objects one wants to amplify.
A large number of methods for image denoising have been proposed in the lit-
erature. Only a few can be mentioned here. For example, Malladi and Sethian [150]
proposed a level-set method that moves the iso-intensity lines of the image's gray
level mountains according to their curvature. The method smoothes contours in a
hierarchical fashion. A min/max-flow criterion is used to stop the algorithm. When
applied to handwriting, only the larger strokes survive smoothing. They are bounded
by sharp edges.
Hierarchical image decompositions using wavelets have been successfully ap-
plied to image denoising. Examples are the systems described by Simoncelli and
Adelson [213] and by Donoho and Johnstone [56]. They transform the image into
a multiscale representation and use the statistics of the coefficients of this repre-
sentation to threshold them. The back-transformed images are then less noisy. This
method works well when the wavelets used match the structure of typical image
details, such as edges or lines. What is problematic with these approaches is that