A first approach to granular metrics is to use the range space, Y, and assume that it has an underlying metric. For , let provides a modification of g so that combines values of g(y'), , in some appropriate way. The set is a sphere obtained by the underlying metric.
A particurlar choice for might be
the mean value of the neighbourhood. This is equivalent of low-pass filtering g, which means that high frequencies will be supressed in the metric. This in turn means that we obtain a metric that is relatively robust to noise.
An other choice of is
Taking the supremum of the neighbourhood might in some applications be motivated. This is mainly true if our has a very good locality, that is, it concentrates to some points in Y. In these cases it is not desirable to smear these points by using a mean value.
We can now define a metric on F(X,L) through
We will use to mean . A normalisation of c can be given as
Constructions in [3] correspond to the use of , where we use .
For natural reasons, is limited downwards by 0 and
since is the same as the supremum [4]. Note that this is true regardless of the used.
Similarly, is limited upwards by and
Since , will be constant over the whole Y.
Suppose we have two images, and , that are transformed to and , see Image . The mean value- is used in this example.