Object Oriented Motion-Segmentation for Video-Compression in theCNN-UM

  • Authors:

  • Tamás Szirányi;Károly László;László Czúni;Francesco Ziliani

  • Affiliations:

  • Analogical and Neural Computing Laboratory, Comp. & Automation Inst., Hungarian Academy of Sciences, H-1111 Budapest, Kende u. 13-17, Hungary;Analogical and Neural Computing Laboratory, Comp. & Automation Inst., Hungarian Academy of Sciences, H-1111 Budapest, Kende u. 13-17, Hungary;Department of Image Processing and Neurocomputing, University of Veszprém, H-8200 Veszprém, Egyetem u. 10, Hungary;Signal Processing Laboratory, Swiss Federal Institute of Technology, CH-1015 Lausanne, Switzerland

  • Venue:
  • Journal of VLSI Signal Processing Systems - Special issue on spatiotemporal signal processing with analog CNN visual microprocessors
  • Year:
  • 1999

Quantified Score

Hi-index 0.00

Visualization

Abstract

Object-oriented motion segmentation is a basic step of theeffective coding of image-series. Following the MPEG-4 standard weshould define such objects. In this paper, a fully parallel andlocally connected computation model is described for segmentingframes of image sequences based on spatial and motion information.The first type of the algorithm is called early segmentation. It isbased on spatial information only and aims at providing anover-segmentation of the frame in real-time. Even if the obtainedresults do not minimize the number of regions, it is a good startingpoint for higher level post processing, when the decision on how toregroup regions in object can rely on both spatial and temporalinformation. In the second type of the algorithm stochasticoptimization methods are used to form homogenous dense optical vectorfields which act directly on motion vectors instead of 2D or 3Dmotion parameters. This makes the algorithm simple and less timeconsuming than many other relaxation methods. Then we applymorphological operators to handle disocclusion effects and to map themotion field to the spatial content. Computer simulations of the CNNarchitecture demonstrate the usefulness of our methods. All solutionsin our approach suggest a fully parallel implementation in a newlydeveloped CNN-UM VLSI chip architecture.