Using the recursive form of Hierarchical Step-Wise Optimization (RHSWO), processing window artifacts are often observed in the segmentation results. Processing window artifacts are region boundaries that are along processing window seams, regardless of the fact that pixels across the seams are very similar. A minor processing window artifact can be viewed a squared off or straight region region boundary, but the noticeability of it depends on the image. Nevertheless, all processing window artifacts can be eliminated by inserting the following step to the definition of rhswo(level,X) given on the RHSWO page:
4. If level = rnb_levels exit. Otherwise, switch the region assignment of selected pixels.
The proccessing window elimination step (step 4 above) includes two approaches to identifying regions that may contain pixels whose region assignment should be switched to that of another region. The first approach analyzes data points on the seam between the data subsections to find pairs of regions that contain pixels that are more similar to the other region. This approach is the most effective at eliminating obvious blocky artifacts at the processing window seam. The second approach compares all of the regions to each other directly, and regions that are relatively similar to each other are assumed to contain pixels that may be more similar to he other region. This approach is most effective at eliminating more diffuse processing window artifacts. Once pairs of regions are identified that may contain pixels whose region assignment should be switched, the pixels in these regions are examined and switched as appropriate.
(a)
(b)
Fig. 6. (a) Region mean image for the segmentation result produced by RHSWO with processing window artifact elimination and BSMSE dissimilarity function at 34 regions and global dissimilarity value of 10.2644. (b) Hierarchical boundary map for (a).
This approach to eliminating processing window artifacts is very efficient because it focuses specifically on the pixels and regions that are involved in producing the processing window artifacts. In addition, performing this step after performing region merging in step 3 of rhswo(level,X) avoids switching pixels between regions that are merged anyways in step 3, enhances the processing efficiency. An additional side effect of this approach to processing window artifact elimination is the reduction in the feature variance or spread of the region clusters, which enhances region quality.