Adaptive filtering of distorted displacement vector fields using artificial neural networks

In this paper the utilization of artificial neural networks (ANN) for motion estimation is considered. By means of simple neural structures it is possible to improve the reliability and accuracy of block matching algorithms (BMA) by a postprocessing of the similarity criterion. An associative memory realizes an adaptive choice of these filtering structures depending on the image contents. The fundamental idea and some results will be described. The performance capability of the proposed method is shown for selected two-dimensional measuring situations which are not solvable with conventional BMA.