Two stage motion-based region merging algorithm for content-based coding

We propose a two stage motion-based region merging algorithm for content-based coding. It uses a six-parameter motion model and the change image obtained from the conventional change detector. First, the type of each region is decided as one of an object, background and uncertain region using the change image. In the first region merging stage, all pairs of two adjacent object regions and two adjacent regions are extracted. Affine motion model parameters for each pair are computed. Then, the merging for one pair that has the smallest error is performed for each type. In particular a large threshold value is used to efficiently merge many redundant regions segmented in the background. In the second region merging stage, the merging for the adjacent uncertain regions or adjacent regions of different types are performed with the same procedure as that of the first stage. A small threshold value is used to merge the extracted adjacent regions which are represented with homogeneous motion parameters while preventing the merging of adjacent object region and background region in the wide range of the threshold value. From the experimental results the proposed merging algorithm is shown to efficiently reduce many redundant regions into one or a few regions in the wide range of the threshold value.