Below is a brief description of the developed software, the list in alphabetical order.
AANN (Autoassociative Artificial Neural Network)
Autoassociative Neural Network simulation system. Neural network consists of input, output and three hidden layers. Number of output neurons is equal to number of input neurons. Intermediate hidden layer is the bottleneck - it has less neurons than input layer. This neural network is able to decrease the redundancy of processing data, validate corrupted data and fill up omitted data. A lot of methods can be used for network training: Back-Propagation, Back-Propagation with inertia term and one-dimensional optimization, usual delta-bar-delta and modified delta-bar-delta method with exponential adaptation, Quick-Propagation, conjugate gradients methods in Fletcher-Reeves and Polak-Ribiere modifications, quasi-Newton method in Broyden-Fletcher-Goldfarb-Shanno modification.
ADF (Adaptive Focusing)
Computer program for reconstruction of defocused and smoothed images. Two algorithms are used: the solution of the system of integral Fredholm equations of the first kind with Tikhonov regularization, and the reconstruction by neural network of special architecture. There is also a possibility in the program to restore a kernel of transformation if initial and distorted images are known. This kernel can be used for reconstruction of other images distorted by this optical system.
ApprPade (Approximation Pade)
Software for Pade-approximation by artificial neural network.
ART2 (Adaptive Resonance Theory 2)
Adaptive Resonance Theory 2 simulation system. ART is the neural network that allows the growth of elements in the training stage. It allows classifying of arbitrary data and can be used for vector quantization, data contrasting and multi-dimensional approximation.
Software for producing stereo films. As input data it uses two films received from pair of synchronized cameras (in AVI format). AVI2 can produce the stereo film for various stereo display devices: anaglyphic glasses, parallax barriers, microlens devices and holographic stereo-screen.
Computer program for the synthesis of static stereo images for microlens devices and parallax barriers with vertical and horizontal orientation.
Three-dimensional scene reconstruction from stereo pair images.
Investigation of arbitrary data strucure and multidimensional data visualization.
DDT (Digital Data Testing)
Software that allows the studying of fractal data dimension and reducing data redundancy via principal component analysis. When one has a huge number of data, fractal dimension is defined by the Box-Counting algorithm while in the case of poor dataset it is defined by the original algorithm. Principal component analysis can be performed by both Oja and Sanger methods.
FAN (Feed-forward Artificial Neural Network)
Feed-forward artificial neural network simulation system. Network consists of input, output and one or two hidden layers of neurons which can have various activation functions. A lot of methods can be used for network training: Back-Propagation, Back-Propagation with inertia term and one-dimensional optimization, usual delta-bar-delta and modified delta-bar-delta method with exponential adaptation, Quick-Propagation, conjugate gradients methods in Fletcher-Reeves and Polak-Ribiere modifications, quasi-Newton method in Broyden-Fletcher-Goldfarb-Shanno modification.
Special software that allows receiving information in real time from CQG and storing it in compact database. This program is resident one - it makes DDE requests to CQG in definite time moments, gets and process necessary financial information and refreshes database.
Computer program that is able to minimize closed loops in directed weighted graph.
Software that utilize classical algorithms of Hidden Markov Models.
HON (High Order neural Network)
Simulating system for modeling of the neural network of high order - up to third.
KOH (Kohonen maps)
Software for Kohonen maps realization.
DLL utilizing standard numerical algorithms in C++ including:
I. Linear algebra methods:
Solution of linear algebraic equations by Gauss elimination with pivoting.
LU matrix decomposition.
Solution of linear algebraic equations based on LU decomposition.
Cholesky decomposition of symmetric positive definite matrix.
Calculation of matrix determinant
Inversion of a matrix by various methods.
Calculation of eigenvalues and eigenvectors of arbitrary matrix.
II. Discrete Transformations:
Various modifications of Fast Fourier Transform
Fast Walsh Transform.
Discrete wavelet Haar transforms.
Linear prediction coefficients calculation (covariance method).
Linear prediction coefficients calculation -autocorrelation Durbin-Levinson method.
Calculation of reflection coefficients (RAPCOR).
Cepstral coefficients calculation.
Various modifications of one-dimensional optimization: golden section, parabolas method and so on.
Various modification of gradient descent including momentum terms and one-dimensional optimization in anti-gradient direction.
Method «delta-bar-delta» - adaptive step of optimization for every variable.
Modified «delta-bar-delta» with exponential adaptation of parameters.
Conjugate gradients methods in Fletcher-Reeves and Polak-Ribiere modifications.
Quasi-Newton method in Broyden-Fletcher-Goldfarb-Shanno modification
Software for estimation of the cost of city land. As initial data it uses the cost estimates of some support lands with known coordinates. MKZ allows: a multi-dimensional approximation of the land cost for all the city territory; data verification and separation of zones with similar cost. The underlying algorithms are based on artificial neural networks: Autoassociative (for data validation), adaptive resonance theory (ART) neural network and radial basis network. Estimation results are mapped onto town map in the form of level lines or area filling. It is possible to export results via DXF format into other CAD systems such as 3D MAX, Intergraph etc. A version integrated in MapInfo exists.
MMC (MultiMedia Classes)
Set of C++ classes for processing of images and sound, including:
CDibBitMap - class for processing of file in BMP (DIB) format.
CPic - class for numerical calculations results visualization both in 2D and 3D manner.
CAviDib - class for processing movie data in AVI format.
CWaves - class for processing sound data in WAV format.
MovieRec (Movie Reconstruction)
Special software for old film reconstruction. It contains a lot of digital filters for correcting the image defects, equalizing light level, improving contrast range and focusing images. Set of selected transformations is assigned to the group of frames manually and all the film or only its part is processed in a batch mode.
NSD (Neural Stereo Display).
Computer system for calculation of images for neuro stereo display.
OCBD (Optical Characters BD)
Software for the creation of the database of scanned characters. This database can be used for training and testing of optical characters recognition (OCR) systems. As an input it uses BMP file that contains the scanned page of text. The user frames every character by computer mouse in a window and fills in attribute form in popup window. After that the selected character is transformed into binary or grayscale representation in a window of 16x16 or 32x32 and added to database together with its attributes.
Simulation system of radial basis function artificial neural network. As basis functions it uses spherical or elliptical Gaussians or parabolic functions. All network parameters - network weights, radial functions positions and metric - are adjustable. A lot of methods can be used for network training: Back-Propagation, Back-Propagation with inertia term and one-dimensional optimization, usual delta-bar-delta and modified delta-bar-delta method with exponential adaptation, Quick-Propagation, conjugate gradients methods in Fletcher-Reeves and Polak-Ribiere modifications, quasi-Newton method in Broyden-Fletcher-Goldfarb-Shanno modification.
Computer program for restoration of intermediate views. As an input it uses stereo pair - i.e. two views of the same scene from shifted cameras. Matching of pixels in both images is established as a result of analysis of gradients and density of intensity. Viterbi algorithm allows then to restore view from arbitrary point between cameras. This software can be used: for transmitting of multi-view images through usual channels - it is enough to transmit only two extreme images; for reconstruction of additional views after aerial photography; for embedding of additional frames into old movies which have low frame-ratio.
Computer program for constructing the "skeletons" of scanned images. It allows to carry out the vectorization of images. The results of images processing can be exported to AutoCAD with the aid of DXF format.
SndLab (Sound Laboratory)
Software for sounds visualization. It allows drawing of arbitrary parts of sound wave in any scale, Fourier and Walsh Transforms and sonograms.
Computer program for producing the characteristic features of arbitrary image. These features are invariant to shifts, rotations and scale transformations.
TS (Time Series)
Adaptive system for multi-dimensional time series forecasting and the construction of the trading systems for traders.
Computer program for recording the dictionaries as a set of wave files. It allows to record, listen and correct wave files and the set of attributes for every word as well. The program can be used as a tool for scientists that deal with speech recognition and voice identification.
VDP (Voice Data Processing)
Computer program for transforming speech into set of characteristic features. As initial information it uses the files in wave format. Output results are the linear prediction coefficients (LPC) and their time derivatives, reflection coefficients (RAPCOR) and their time derivatives and cepstral coefficients and their time derivatives.