Simulated annealing split and merge segmentation algorithm for object based video codecs with the output bit rate as the objective function

This paper proposes a strategy for encoding an image into a database structure which holds all pertinent information about the image, namely the spatial regional location of objects, their statistical representations and certain texture classifications. These are then operated upon by a simulated annealing algorithm to make a decision as to which objects should be merged to generate the smallest number of possibly important areas. A neural network approach is then presented for applying this knowledge to a sequence of video images and of tracking these objects within the sequence