Neural Abstraction Pyramid Architecture - Hierarchical Neural Networks for Image Interpretation - page 81

Fig. 4.9. Local contrast normalization. View of layer l = 2 . The contrast C ± l is divided by

the local contrast level S l . See text for a more detailed explanation.

The normalized image was computed by a Neural Abstraction Pyramid that was
iterated 15 times. The network consists of five layers with resolutions from 320 × 240
decreasing to 20 × 15. It computes subsampled versions G l of the high-resolution
input G 0 . In each layer, local contrast is detected using 5 × 5 center-surround kernels
and divided by the smoothed squared contrast. The normalized contrast is combined
in a top-down fashion.
Figure 4.9 illustrates the iterative operation of a single network layer. In the
following, the templates used for the computation of the different features are de-
scribed in detail. The feature arrays are updated in the listed order. If not stated
otherwise, the forward and lateral projections are direct and the backward projec-
tions are buffered. The projections have only one input from the offset (0 , 0) with
the weight one and are summed by the output unit. The units have linear transfer
functions and zero bias. The activity of the cells is initialized to zero. The intervals
indicate the scaling of activities used in the figure.
• G l - intensity [0 , 1]
- contains shrunken versions of the original image
- has only a single forward projection that averages 2 × 2 windows of G l− 1
• C ±

l - contrast [0 , 1]
- contains local contrast, separated by sign
- lateral projection has center-surround input from G l (DoG 5 × 5 - 3 × 3) and linear
threshold transfer function f lt
- buffered lateral projection receives input from corresponding cell in D

±
l