Improving dynamic programming to solve image registration

Abstract

A new method of dynamic programming (DP) is presented. A picture is paired with a reference map. In order to arrive at picture registration, an ordered sequence of primitive patterns is formed from the map. An attempt is made to locate a similar sequence from the picture using classical picture processing methods. This yields a distorted and unordered sequence marked by noise. DP is then used to search for a common match to the two sequences. Both the initial and final registration states are, however, unknown. Moreover, the continuity of the path is frequently interrupted, due to shortcomings in detection, or perhaps occlusions. Thus, classical DP algorithms appear to give only poor results under practical conditions. When well-adapted to treat signals issued from a first order Markovian process, they can be used, as this method shows, to treat kth order processes. The concept of the “virtual state”, kept in memory, suspends broken paths until they can be reconnected, thus bridging the existing gaps between the different parts of the path. The proposed algorithm has been applied to satellite images with their contours sectioned by clouds. The result leads to the guarantee of very good registration, in spite of image contours badly hampered by noise and/or clouds.