Home – Home – Jornals – Avtometriya 2022 number 5

The paper presents a new robust video watermarking method. Its main idea consists in adding temporal sinusoidal sequences to each pixel of the video signal. The two-dimensional field of their phases corresponds to the watermark image. Simple and fast algorithms for embedding and extracting watermarks are described. The results of experimental studies demonstrate the high watermark extraction quality, robustness to some attacks, including temporal desynchronization, and high visual quality of the protected video.

In a space of tree-structured object representations, the accuracy of object classification in terms of an error probability depending on the amount of processed information is studied. For a given set of objects, the lower bound to the average error probability as a function of the average mutual information between the objects and the decisions about their classes is given. Using multilevel discriminant functions in the set of object representations, a guided search algorithm for an object class-label decision is proposed, and a computational profit of the guided search relative to the exhaustive search is shown analytically. In the source datasets of face and signature objects given by the grayscale images as well as in an ensemble of these datasets, we calculate the experimental dependences of the average error probability and the average mutual information on the algorithm parameter which defines the above-mentioned computational profit. Also, for both source datasets and their ensemble, we give numerical values of the lower bounds to the error probability that allow us to estimate the redundancy of the algorithm error probability for different values of the computational profit.

The goal of registering point clouds in a 3D space is to find an orthogonal transformation that maximizes the consistent overlap of two point clouds. The most common registration method using purely geometric characteristics is the Iterative Closest Points (ICP) algorithm. The disadvantage of the classical ICP variants is the dependence on the initial location of the point clouds. Coarse registration algorithms are used to find a suitable initial registration of two clouds. In this paper, we propose a new algorithm for extracting common parts and coarse registration of point clouds.

In this paper, we propose a new two-level method for combining image segmentation maps based on minimizing the two-objective quality functional. The functional is formed as a weighted sum of the information redundancy measure and variation of information computed from the original image and the combined segmentation map. Applying such a measure, we obtain an image partition that provides a compromise between the objectives of minimizing the number of outlined informationally important segments and minimizing the information difference between the original image and the resulting partition. The proposed method improves the result of segmentation in comparison with the method for combining partitions based on the criterion of the minimum information redundancy.

The article deals with the issues of image restoration when only a part of observations subjected to additive noise regularly placed in the original image is available. In other words, the problem of restoration of a thinned image is solved (based on pilot pixels). Pixels themselves are mixed with the white Gaussian noise. To solve this problem, special modifications of nonlinear filters are synthesized based on deep doubly stochastic Gaussian models. The results obtained allow us to draw a conclusion about the effectiveness of the proposed filters in comparison with linear methods and traditional algorithms. The study shows that images can be reconstructed based on only 50% of information using a doubly stochastic model, resulting in a relative error of only 9%.

A technique for eliminating the influence of the image interpolation effect in the algorithm of estimating image spatial deformations is proposed. The technique is considered for similarity measures of images used in the synthesis of algorithms, in particular, the mean square of the inter-frame difference, the inter-frame correlation coefficient, and the Shannon mutual information. Calculated expressions for compensating the influence of bilinear and bicubic interpolation are obtained. The developed technique is also applicable to other similarity measures used in the development of algorithms for estimating spatial deformations of images, as well as any interpolations: spline, using Lagrange and Newton polynomials, power functions, etc.

The technology of constructing a recursive filter on a non-uniform grid of samples with parameter identification on test images is discussed. This filter is the IIR-filter that has a physical feasibility problem. To overcome it, a multi-step procedure is implemented. Unfortunately, identifying the best filter in terms of a given criterion does not guarantee that a recursive implementation of that filter will be stable. In the paper, for the considered iterative scheme, stability conditions are obtained. It has been experimentally confirmed that if, these conditions are met, it is possible to achieve a high quality of the correction. Based on the obtained criteria, a technology for correcting defocusing with control over the stability of estimates is proposed. The results of image correction showing the effectiveness of the technology are presented.

The article explores the reliability of training samples used to train convolutional neural networks for the diagnosis of pulmonary diseases. It is shown that the sample, in which 3500 X-ray pictures of healthy patients and the same number of pictures of patients with tuberculosis, is very heterogeneous. When training on different parts of the sample and recognizing its various parts, significantly different results are obtained.

An image compression technology based on machine learning is developed. Segmentation of the original image into discarded and stored zones is applied. An algorithm of compression of stored zones based on the nested coverage of the image is used. Discarded zones are replaced by a reliable fake during decompression. Machine-learning algorithms based on autoencoders, convolutional and adversarial neural networks are used at all stages of compression technology (segmentation, pixel approximation of stored zones, fake of discarded zones, etc.). Computational experiments are performed to study the proposed compression technology and the included machine learning algorithms in natural images. The results of computational experiments confirm the prospects of the proposed technology for problems related to digital image compression.

A two-stage method for adaptive traffic signal control based on an estimate of the predicted weighted traffic flow passing through an intersection is proposed. At the first stage, we estimate the travel time required for each vehicle to pass the intersection using an artificial neural network model and estimate the predicted traffic flow through the intersection for a given phase of the traffic signal cycle. At the second step, a weighted flow estimate is formed, which takes into account the waiting time of vehicles. The proposed method for choosing the traffic signal phase is based on maximizing the weighted traffic flow. The results of experimental studies allow us to conclude that the proposed approach outperforms the classical approaches and state-of-the-art methods of traffic signal control based on reinforcement learning.

The results of numerical simulation based on the Mie theory of the superresonance effect for a dielectric sphere with a low refractive index are presented. Water is used as a material of the mesoscale sphere. It is shown that not only the previously studied weakly dissipative mesoscale spheres made of a material with a “medium” (about 1.5) and high (more than 2) refractive index, but also a low one (about 1.3) support the high-order Fano resonance effect associated with internal Mie modes. In this case, the intensities of resonant peaks for both magnetic and electric fields in the vicinity of the poles of the sphere can reach extremely high values of the order of 10⁶-10⁷ for a water droplet with a Mie size parameter of about 70.

The effect of changes in the thermophysical properties of the molybdenum film during intermediate oxidation during laser ablation on the dimensional effect of track formation is investigated by numerical simulation. In accordance with the data obtained, the hypothesis explaining the reduction of the track width in the ablation zone of the film in comparison with the effective diameter of the laser beam is refined. It is shown that a specific change in the thermal conductivity coefficient of a substance at the time of oxidation has a significant effect on the distribution of the temperature field, expressed in narrowing of the characteristic of the temperature distribution over the film surface, which has not been previously considered. It is established that the change in the density, specific heat capacity, and thermal effect of the chemical reaction of molybdenum oxidation during film oxidation does not significantly affect the temperature distribution in the zone of exposure to the laser beam.

The features of formation of face images of extended absolutely slit objects with an arbitrary opening of holes are studied in a coherent-optical system. On the base of the constructive approximation of the spectra of spatial frequencies (Fraunhofer diffractional patterns) of objects with different dimensions of front and back apertures, formulas for the field in the image of the front face are obtained and analyzed. Cases of objects with significantly expanding and significantly narrowing apertures of an extended hole (the differences in aperture sizes are much larger than the size of the Fresnel zone) are investigated in detail. It is found that the image structure of faces depends on the type of the opening. With a positive opening, where the back face is in the shadow region, the field at the output of the system corresponds to the image of the front face; with a negative opening, the image of the back face is observed. It is shown that error of determination of boundaries in the image of the active face of the object is inversely proportional to the square of the aperture difference. The invariant properties of the projection system for forming images of external faces of extended holes with an arbitrary opening are analyzed.

Design principles for the implementation of an infrared microscope registering the intrinsic thermal radiation of objects are considered. The infrared microscope temperature resolution, depending on the optical system parameters, design and photoelectric parameters of multielement infrared focal plane arrays (IR FPAs) are analyzed.