Kari-Pekka Estola

Pulse pattern excited linear prediction voice coder

Speech coding of the code excited linear predictive type is implemented by providing an excitation vector which comprises a set of a pre-determined number of pulse patterns from a codebook of P pulse patterns, which have a selected orientation and a pre-determined delay with respect to the starting point of the excitation vector. This requires modest computational power and a small memory space, which allows it to be implemented in one signal processor.

PDF estimation using order statistic filter bank

Introduces a new probability density function (pdf) estimation method based on the concept of order statistic filter banks. The proposed method offers a new kind of possibility for statistical analysis of signals with real time processing capabilities. Besides the new real time processing method offers a procedure to control such properties of the pdf estimate as the variance and the resolution.<<ETX>>

Designing power-of-two-Nth band FIR filters using Remez algorithm

The author introduces a novel method for designing Nth-band filters for cases where N is an integer power of two. The proposed design method exploits the multiple exchange Remez algorithm. Thus, the design method is easy to implement and the design process is fast. The novel design method is capable of designing nearly optimal Nth-band filters with weights both on the passband and the stopband. The proposed filter structure is similar to the structure of halfband filters. Time domain specifications for the resulting Nth-band filters are fulfilled structurally.<<ETX>>

New class of order statistic filters for running median estimation

A novel low variance median estimator is presented. This order statistic based estimator is derived by means of a set-theoretic approach. It is shown that the proposed subset average median estimate (SAME) shares many good properties with both the mean and the median operator. Properties of this filter are controlled with two parameters: the window length and the subset size q. A good noise attenuation together with robust behavior can be obtained by selecting an appropriate subset length q. Numerical studies show that SAME filter outperforms the ordinary median filter in noise attenuation when is is applied to signals with various noise characteristics including Laplacian and Gaussian noise.<<ETX>>

Design of computationally efficient FIR filters for sampling rate alteration and multiband filtering with arbitrary passbands and time response

A general method is presented for designing computationally efficient FIR filters for sampling rate alteration and multiband filtering with arbitrary frequency response and time response. These filters are constructed by cascading subfilters whose delays are integer powers of the unit delay. The non-linear approximation problem is solved using the dual simplex algorithm enabling simultaneous use of multiple sets of constraints including the time response constraints. The basic algorithm is derived from the general framework for minimax solution of overdetermined systems of non-linear equations.

Hybrid multivariate marginal median filters

This paper introduced a new class of hybrid multivariate marginal median filters. The proposed approach combined the noise suppression capabilities of the componentwise median filters and the edge preservation properties of the vector median filters. The proposed filter class was based an the use of multiple marginal orderings of the input data set. This assured that the optimal projections were included in the computations. It was shown in the examples that the proposed filters outperform other median filters in random noise suppression and in the preservation of the shapes.<<ETX>>

Subset averaged median estimators

Alternative interpretations and realizations for the subset averaged median estimators are discussed. Maximum noise suppression performance of the new filter class is achieved using the order statistic (OS)-filter realization. An alternative structure is obtained realizing the new filters directly in the sample domain. This leads to highly parallel realization of the new filters. The time domain interpretation of the subset averaging method allows adjustment of the time domain properties of the filters. The time domain weighting of the data within the window can be achieved using the samples to be weighted more often in the subsets than other samples. Midpoint weighting preserves the edges exactly, whereas uniform weighting results in gradual transition.<<ETX>>

Adaptive nonlinear image restoration filter

A new adaptive nonlinear image restoration filter structure which has both good noise rejection and excellent ability to preserve sharp edges is proposed. The good performance of the proposed filter is based on the adaptation of the window size and the new near edge correction (NEC) method. By utilizing the enlarged window is stationary areas, high noise rejection is achieved. Moreover, small windows during transient events together with the NEC method guarantees the good preservation of sharp edges also when image is corrupted by additive wideband noise.<<ETX>>

High performance varible cutoff frequency linear phase filters

An approach to realizing varible-cutoff-frequency linear-phase filters is presented. Selective and computationally efficient filter structures previously developed by the author (Proc. IEEE Symp. Circuits and Systems, p.305-8, 1988) are used. It is shown how the coefficients of these filters can be fairly easily varied to obtain computationally efficient transformed filters. The calculation of the updated coefficients can be done in real-time. The resulting filters are efficient for cases where the transition band is narrow. This is because the proposed filters effectively realize pole pairs in the transition region. In this approach, the poles outside the unit circle are approximated with zeros.<<ETX>>

Interpolated Gaussian scale-space filters

Computationally efficient interpolated Gaussian scale-space filters are introduced. The proposed scale-space filters incorporate multirate signal-processing methods and are realized using multistage filter structures. The scale-space filters can also be used in decimating the data, if desired. The proposed filtering methods require dramatically less computation than conventional Gaussian scale-space filtering, especially where the scale changes with an integer factor. The filters are extremely efficient when the change in scale is an integer power of an integer constant.<<ETX>>