Variations of Markov - Varying Markov Mask

The mask in Gauss Markov Random Field models defines the input space to the model and hence the algorithm's field-of-vision in the image. We consider here two masks; each producing the same number of features but with a different field-of-vision.

The two masks are the standard fourth-order symmetric mask and a cross mask, shown below. If is obvious that the cross mask has maximum field-of-vision in the four orthogonal directions but has no "vision" in any other direction. The fourth-order mask, on the other hand, has limited "vision" in all directions but maximum localised "vision".

The results in the table below indicate a definite superiority of the fourth-order mask for tests involving microtextures. It also shows a marked advantage of the cross mask for macrotexture tests. However, it should be noted that the macrotextures in these groups are brick images with mortar at 0/90 degree angles. The performance difference may not be so distinct when these angles are not significant. Some hint of this can be seen when comparing the disparity between the results for the different masks on the bomb and bombRot test suites.