Binarization of Matrix Codes - Hierarchical Neural Networks for Image Interpretation - page 163

Fig. 8.11. Recall of network trained for binarization of Data Matrix codes. The development
of all feature array activities is shown for one of the degraded test examples.
few iterations, the non-uniform background level causes the upper part of the code
to have higher activity than the lower part. This inhomogeneity is removed during
refinement. Furthermore, the output is driven from intermediate gray values that
signal uncertainty towards black and white, which is characteristic of the desired
output.
In the lower row of the figure, the only hidden feature array of the lowest layer
is displayed. Here, a representation of the cell structure emerges. Bright areas of
the input image are broken into a discrete number of cells. For each bright cell, an
activity blob rises and remains stable. Adjacent blobs are connected either vertically