Chair: John Spivy, Purdue University, USA
D. Docampo, Universidade de Vigo (Spain)
S. R. Baldomir, Universidade de Vigo (Spain)
This paper presents computationally efficient algorithms for function approximation with hinging hyperplanes. Approximant units are added one at a time using the method of fitting the residual. To fit an individual unit we have to solve a sequence of Quadratic Programming problems, an approach which has proven to offer significant advantages over derivative-based search algorithms. Empirical results on synthetic data illustrate the main characteristics of the algorithms.
Rusen Oktem, Tampere University of Technology (Finland)
Moncef Gabbouj, Tampere University of Technology (Finland)
A new multiresolution decomposition scheme based on constrained B-Spline decimation/interpolation is proposed. The proposed decomposition algorithm is tested on greyscale and color image compression, and the results are compared with DCT based compression. Dithering and postprocessing are applied to the compressed images to deal with artifacts at low bit rate compression.
Karin Vibe-Rheymer, Swiss Federal Institute of Technology (Switzerland)
Jens Timmer, Albert-Ludwig-Universitaet (Germany)
Jean-Marc Vesin, Swiss Federal Institute of Technology (Switzerland)
Fractal signals have attracted a lot of attention in various fields lately, and numerous algorithms have been designed to analyze them. Most of them investigate long-term correlations, hence requiring long data sets (i.e. data sets extending to very large time scales). This requirement is however rarely met in practice, which can cast doubts on the reliability of the results. This work tries to partially fill this void by analyzing a method examining long-term correlations for short time series. It is shown the conclusions obtained for long data sets remain valid, but there are some particular cases that should be taken into account before concluding on the fractality of a signal. A practical example, namely heart rate recordings, is taken to illustrate some possible pitfalls that can be encountered when real-world short data sets are to be studied.
Paolo Palazzari, ENEA - HPCN Project (Italy)
Moreno Coli, University "La Sapienza" (Italy)
In the last years Image Fractal Compression techniques (IFS) have gained ever more interest because of their capability to achieve high compression ratios while maintaining very good quality for the reconstructed image. The main drawback of such techniques is the very high computing time needed to determine the compressed code. In this paper, after a brief description of the IFS theory, we discuss its parallel implementation by comparing the different level of exploitable parallelism. In the paper we show that Massively Parallel Processing on SIMD machines is the best way to use all the large granularity parallelism present in this problem. Finally, we give some results achieved implementing IFS compression technique on the MPP APE100/Quadrics machine.
Ignacio Santamaria-Caballero, University of Cantabria (Spain)
Carlos Pantaleon-Prieto, University of Cantabria (Spain)
Anibal R. Figueiras-Vidal, Universidad Carlos III de Madrid (Spain)
This paper presents a method for off-line segmentation and AR modeling of signals characterized by abrupt changes between stationary segments (quasistationary signals). Assuming that the number of models and their orders are known, we propose a suboptimal procedure for maximizing the likelihood function based on the Expectation-Maximization algorithm. At each iteration the transitions are estimated as the posterior probabilities that a sample was generated by a given model (E-step); then, the new set of models is obtained by solving a least squares problem (M-step). It is shown by means of computer simulations that the algorithm achieves accurate estimates of the transitions and AR coefficients with a moderate computational complexity.
Sari Siren, Tampere University of Technology (Finland)
Pauli Kuosmanen, Tampere University of Technology (Finland)
Differentiation of a signal is required in many applications in the field of signal processing. Good linear differentiators exist which can be used also in the noisy conditions if the noise is not impulsive. In this paper we consider noise corrupted discrete-time measurements whose time derivatives we estimate. We have experimentally evaluated various nonlinear methods and according to our results any of them is superior to the others. In this paper we propose two methods whose performance has been satisfactory in our experiments for differentiating a signal simultaneously corrupted by Gaussian and impulsive type of noise. First of the methods is median prefiltering followed by linear FIR dilferentiator and the second method is based on robust regression. We also address the problem of second order differentiation.