The performance of the proposed metrics is highly dependent on what transformation that is used. In this case, we used a a spatial gray level dependency (SGLD) matrix of the texture when testing the -metric. The matrix is formed with the position operator 'one pixel to the left'. For more information on the forming of the SGLD matrix, see for instance [6].
When testing the -metric we used a somewhat more sophisticated underlying metric. First two SGLD-matrices, and , were calculated for each texture , one with the position operator 'one pixel to the left' and one with the position operator 'one pixel up'. If the texture is mainly horizontal, will be rather centered arround its diagonal, while will not, and vice versa. A measurement of a matrix 'diagonality' is the moment
This was calculated for both and . yields low values for matrices that are centered around their diagonals, e.g see [7]. Hence, it is possible to calculate the main angle of the texture by
And the underlying metric is defined as
To be able to test the metrics, textures with known distances were generated, see Image . The angle of the texture is known since it is generated by a known Markov random field process. The direction of the texture is so that . is measured in degrees. Given two textures with direction and , the reference distance is calculated according to:
Since a metric is a relative measurement it is not very interesting to compare the numerical values of the distances between two textures from different metrics. However, the relative order among textures within a metric is the same regardless of the metric used. That is, if in one metric, for all other metrics.