Home – Home – Jornals – Avtometriya 2020 number 6

Wavelet filtering algorithms are widely used for image processing, and most of these algorithms are of a threshold nature: the wavelet expansion coefficient whose absolute value is smaller than a certain threshold value vanishes; otherwise the coefficient undergoes some (most often nonlinear) transformation, which is defined by the threshold function used. When high-contrast images are filtered, such processing of expansion coefficients can lead to the emergence of the so-called artifacts in the filtered image - in the vicinity of a jump-like change in the amplitude of the filtered image, there appear pulses that are absent in the original image, which significantly degrades the filtering quality. To eliminate such artifacts, several approaches are considered in the present paper, and a method is proposed, which implies post-processing of the results of wavelet filtering of images by nonlinear algorithms of spatial filtering. The computational experiment is performed, which makes it possible to conclude that the proposed method is an effective way to eliminate artifacts of wavelet filtering of high-contrast images.

A modification of the fast algorithm of bandwidth selection of kernel functions in nonparametric probability density estimation of the Rosenblatt - Parzen type is proposed. Fast algorithms for optimizing the kernel density estimates can significantly reduce the time costs when selecting their bandwidth, as compared to the traditional approach. This is especially true when processing large volumes of statistical data. The basis of the proposed method is the analysis of the formula for the optimal calculation of the bandwidths of kernel functions and the detected dependence between the nonlinear functional of the second derivative of the reconstructed probability density and the antikurtosis coefficient. The proposed algorithm for bandwidth selection provides reduction in the probability density approximation error as compared to the traditional approach. The conclusions from the study are confirmed by the results of computational experiments. Particular attention is paid to the dependence of these properties on the amount of initial information.

A method for reconstructing functionally defined surfaces from stereo images of real objects is proposed. A well-known approach to calculating volume models based on images of silhouettes in the object space is used at the stage of preprocessing. A well-known area-based algorithm with correlation of image intensity levels is applied to calculate the depth map. In order to reconstruct a smooth surface with a compact description, a new method of transformation into a functionally defined model is proposed.

An algorithm for dynamic clustering of temperature signals, which is used to solve the problem of isolating climatic regions on the Earth surface characterized by a certain homogeneous (within the boundaries of the region) type of the climate is considered. A numerical experiment implemented to assess its stability is described. The reference signal model used in the numerical experiment is represented as a sum of harmonic components. The additive noise components are generated in the frequency domain. The parameters varied in the numerical experiment are the noise value and sample size. The estimates of the root-mean-square deviations and stability of the dynamic clustering algorithm are given.

This study is aimed at a comparative analysis of the developed classical machine learning models based on linear models and decision trees, as well as modern algorithms of convolutional neural networks and a neural network autoencoder for solving the problem of predictive detection of pre-failure and emergency conditions of aircraft engines. The simulations are performed using a NASA dataset based on sensor data from aircraft engine life cycles. Several formulations of problems are considered: problems of binary and multi-class classification of normal, pre-failure, and emergency states of aircraft engines, a regression problem for predicting the exact number of operating cycles before the engine failure, and an unsupervised learning problem in which a neural network autoencoder is used to detect abnormal operating cycles of an aircraft engine. The resulting algorithms are combined into a programming framework, which can be useful for analyzing a wide range of predictive maintenance data.

High-speed algorithms for detection and localization of randomly distributed pulse-point objects capable of generating instantaneous delta pulses at random times are investigated. The search is carried out using a receiving device (detector) that can freely move within the search interval and dynamically alter the scanning window size. In this study, a priori information on the distribution of the sought signal source is limited to single-mode functions with a stepwise probability distribution density, which ensures physical feasibility of the algorithms. The parameters of the optimal search are calculated depending on the a priori distribution density of the sought signal source and the required localization accuracy.

A modification of the digital holographic interferometry method for determining the deformations of an object with a diffuse surface by comparing the complex wave fronts reflected from it in two different states is discussed. The difference between the proposed method and the classical methods of digital holography lies in the fact that digital holograms used for image reconstruction have complex values determined by the method of phase shifts (“complex” hologram), while the classic digital holograms have only real values. In addition, the reconstruction of digital holograms is performed with due allowance for inhomogeneity of the reference beam intensity, which increases the quality of reconstruction. For this purpose, a tunable neutral optical filter is introduced into the reference arm of the optical scheme of the interferometer. An experimental verification of the method is performed.

A method of objects image reconstruction in sections by means of a ring antenna array with space sounding by an ultrasonic monochromatic wave is proposed. The image reconstruction is carried out by scanning the internal structure of the object placed inside the ring controlled by the focus of the ring antenna array. A control algorithm for phase shifters in channels of elementary radiators of the ring antenna array for changing the focus position is described. Model results of image reconstruction of elementary point objects are presented. The problems of ring antenna array implementation are discussed.

As applied to dimensional inspection of 3D objects, specific features of the formation images of a volumetric asymmetric absolutely absorbing edge, which is the main fragment of extended (in depth) plates of constant thickness, in the diffraction-limited projection system of images are analytically investigated. The structures and intensity profiles in images of the front and rear faces of the object are studied, respectively, with small and large apertures of the optical system for various relationships of the object bevel c , Fresnel zone size d_d ~ ( l is the light wavelength and d is the object thickness), and angular aperture of the optical system . For the bevel , at a large depth of focus of the system, there is a shift in the intensity profile in the 3D edge images, which is proportional to the Fresnel zone and the bevel value. It is analytically shown that a clear image of either the front face (the rear face is in the shadow) or the rear face (the front face is shaded by the rear face) is observed in the plane conjugated with the front face in the case with large bevels ( ) and small apertures of the system. Formulas for the profile of rear face images are derived and investigated in the case of significant volumetric effects, where the depth of focus of the system is much smaller than the object thickness. The results of the calculations are in good agreement with the results of computer modeling.

The processes of acousto-optical interaction at infra-low frequencies are studied using seismic vibrators as sources of low-frequency acoustic oscillations propagating in the atmosphere and laser measuring lines as optical receivers of vibrations. The proposed types of the source and receiver for the purpose of studying acousto-optical interaction determine the novelty and originality of the proposed approach. The results of experiments on evaluating the characteristics of acousto-optical interaction at infra-low frequencies in problems of laser environmental monitoring are presented.

In this paper, an algorithm for 6-coordinate moving object tracking on a sequence of RGB images based on a convolutional neural network is proposed. Network training is performed on the synthesized data of an object with a dynamic motion model. A Kalman filter is included into the feedback from the output to the input of the network to obtain a smoothed estimate of the object coordinates. Preliminary results of object tracking on synthesized images demonstrate the effectiveness of the proposed approach.

Based on the previously developed numerical analytic method for the description of the input/output parameters of a multidimensional dynamic object for a given set of admissibility (Part I), a method for optimal estimation of the values of continuous linear functionals (numerical characteristics) of the measured parameters is considered in this paper based on incorrect data containing both the fluctuation error and singular noise (Part II). The method ensures the maximum possible decomposition of computational procedures; it does not require performing traditional linearization operations and selecting an initial approximation; moreover, it does not need calculating the spectral ratios in finite linear combinations (with specified basis functions) that describe the cumulative curves of the differential equation, measured parameters, and singular noise. The random and systematic errors are analyzed; an illustrative example and recommended practice for the results obtained are provided.

Designs of a warm (uncooled) aperture diaphragm in a gas-filled cryostat with a matrix photodetector sensitive in the infrared spectral range are studied numerically and experimentally. It is shown that the matrix photodetector with a format of 384 × 288 pixels is cooled with liquid nitrogen to operating temperatures (85 K) in less than 40 seconds if the chosen design of a warm aperture diaphragm is used and in 15 seconds in the case with a throttle microcooler, while maintaining the standard sensitivity of photodetectors based on mercury-cadmium-tellurium (MCT) structures.

A simple way for measuring the radial clearances with controlled accuracy due to reducing random error components is considered. Quantitative estimates of the accuracy, speed, and efficiency of the proposed method are provided. The method is stochastic, is based on the characteristics of the random errors and impeller rotation process periodicity, and makes it possible to perform measurements with the same accuracy over the entire range of rotation speeds of the rotor of a turbo machine. The method considered in the study allows avoiding the procedure of approximation of digital samples.

An algorithm for solving the optimal control problem with terminal constraints is developed. An optimal control problem with terminal constraints and constraints on the control parameter is formulated. A numerical algorithm for solving the problem based on the penalty method and genetic algorithms is described. A computational experiment for the reaction of obtaining phthalic anhydride to achieve the maximum yield of the reaction product in the presence of terminal constraints is performed. The optimal temperature regime and optimal concentrations of reagents were obtained.