This paper considers the problem of detecting surface defects based on a small number of labeled images. The method consists of two stages. At the first stage, all areas that are suspected of containing image defects are found using a selective search algorithm. At the second stage, image recognition is performed in the selected areas based on a Siamese neural network. The method uses multiple images of defects of the same class to train the network. The method is based on training a network to match tasks between similar data. Both images are processed by convolutional subnetworks. The features of both images are calculated. The difference between two feature vectors is calculated. Sigmoid activation is used to determine whether images belong to the same class. In the study of the developed method, the mean average precision indicator was 91.3%, which satisfies industrial applications. The advantage of the presented method is that training occurs on a small set of images. This makes it possible to significantly reduce the costs of data collection and processing.

Formation of features of microstructure images for their subsequent classification often encounters difficulties due to the high variability of images caused by changes in lighting, angle, and scale. Classifying images after feature discovery is also complicated by intraclass variability, where objects of the same class can differ greatly from each other, and interclass similarity, where objects of different classes are similar to each other, which also complicates the task. This paper proposes the formation of image features based on their topological decomposition. The use of the proposed approach makes it possible to avoid the problems described above since the generated features do not depend on the variability associated with changes in the brightness and geometric properties of images. The study was conducted on two datasets: images of metals containing 6 classes and images of polyvinyl alcohol cryogels containing 20 classes. The conducted testing showed that the classification accuracy when using the proposed approach is 94.11% on metals and 80.61% on polyvinyl alcohol cryogels. When using neural networks, the accuracy was 92.16 and 55.45%, respectively. Experiments have shown a preference for using the proposed approach when there is interclass similarity in the dataset.

Face recognition systems have been substantially impacted by the COVID-19 pandemic, as masks cover crucial facial features that renders conventional techniques difficult to implement. A novel methodology for occluded face recognition is presented in the study, based on two-way transfer learning with soft decision fusion. In this approach, pairwise combinations of five pretrained models, namely, ResNet-50, Xception, EfficientNet-B0, DenseNet-121, and MobileNetV2, are used to extract pretrained embeddings of facial images from the 5-Celebrity-Faces dataset that is augmented with a synthetic mask artificially placed on each facial image. These embeddings are passed to three machine learning classifiers, namely, categorical boosting, logistic regression, and linear discriminant analysis. The posterior class probabilities of individual classifiers for pairwise model combinations are fused by averaging to gain highly accurate classification results. The experimental results show that the soft decision fusion of MobileNetV2 and DenseNet-121 outperforms all other models with the maximum accuracy of 91.67% achieved using the linear discriminant analysis classifier.

The paper presents an analysis of the formalism of topological approach to analyzing poorly formalized problems based on the fundamental concepts of functional analysis. The results of the analysis made it possible to formulate a plethora of new approaches to the definition of the lattice estimates and of the ways of introducing metrics on lattices which arise over the topologies of the feature values. In particular, the use of so-called “support chains” to analyze Boolean lattices formed over Zhuravlev-regular sets of precedents allowed here to formulate a cutting-edge research area that consists in replacing the estimates of the lattice elements with certain types of the functions and/or of the vectors. The analysis also allowed to propose several new approaches to a systematic study of semiempirical (tunable) distance functionals known from the literature. These functional are applied as means to generate feature descriptions in the course of solving various applied problems. The analysis of the precedents’ relations between the feature values and the target variable as sets of interactions of Boolean lattice elements indicated the possibility of generating synthetic features using metric distance functions. The paper formulates a few of perspective approaches for (1) estimating the relevance (or “informativeness”) of the metrics in respect to the problems to be solved and for (2) generation/selection of synthetic features, more informative than the initial feature descriptions (that generated the topology and the corresponding lattice). The paper also presents the results of experimental testing the algorithmic approaches based on the formalism developed. The computational experiments were performed with 2400 independent datasets from ProteomicsDB dealing with “molecule-numerical property” type of data. The experiments allowed to produce quite efficient algorithms for predicting numerical properties of the molecules (rank correlation in cross-validation was found to be 0.90 ± 0.23 when averaged over the 2400 datasets). The analysis of the results of experimentation indicated the metrics that most often generate the most informative synthetic features and the forms of corrective operations characterized by the best generalization capacity.

This paper proposes several methods for increasing the consistency of data structure using the example of the problem of searching for groups of identical products on a marketplace. Proper grouping of identical products makes it easier for the buyer to find the most suitable offers in terms of price, rating, or delivery time, and thus improves their user experience. These groups also allow marketplace sellers to gain insight into the competitiveness of their prices and the attractiveness of their product listings compared to other sellers, so they can make changes if necessary. The main objective of the study is to select the optimal strategy for combining products into groups based on the principle of identity. To this end, experiments were conducted with various approaches, from which the most effective one in terms of quality and speed of data processing was selected. The developed method takes into account various characteristics of goods and their interrelations, which allows for the accurate identification of identical goods. Among other things, the chosen approach has an efficient implementation in the MapReduce paradigm, which makes it relatively fast even on large volumes of data. This, in turn, improves user interaction with the platform and increases the overall efficiency of the marketplace.

Lower bounds to an amount of information required to make decisions with a given admissible average error for probabilistic models of discrete source coding over distorted letters, image segmentation, and pattern recognition (object classification) have been studied. The indicated bounds have been constructed as generalizations of the well-known information-theoretic trade-off relation in the form of a rate distortion function with Hamming distortion measure. The proposed bounds are independent on decision algorithms and are useful for evaluating an efficiency of the algorithms in terms of a redundancy of the amount of information used by an algorithm with an acceptable error probability or a redundancy of the algorithm error probability relative to the boundary value at the specified amount of information used.

Numerous studies of human brain activation patterns based on electroencephalogram data show contradictory results. Therefore, the search for and analysis of methods that allow identifying stable correlations of these patterns with the characteristics and states of subjects, their cognitive abilities, and types and schemes of experiments is an extremely urgent task. The present study focuses on the search for features obtained from electroencephalogram data that robust correlate with human arithmetic abilities. The elements of the initial set from which the subset of informative features is extracted are the metrics of the graph constructed using five methods of computing the connectivity matrix, which characterize in different ways the degree of spatial synchronization of electroencephalogram signals (coherence, imaginary coherence, phase lag index, directed phase lag index, weighted phase lag index), which were calculated in five frequency ranges (delta, theta, alpha, beta, gamma). For each method of computing the connectivity matrix and each frequency range, a connectivity graph is constructed, which is determined by pairs of significant interactions between areas of the brain, and then, for each such graph, seven connectivity metrics are calculated, ultimately forming a set of 5 × 5 × 7 = 175 initial features. The studies are repeated for different values of the thresholds of significance of interaction of channels when constructing connectivity graphs. With this issue formulation, the problem of multicollinearity of graph metrics arises, which does not allow obtaining stable results. To solve it, the paper proposes an approach the essence of which is to select a feature with the highest absolute value of the rank partial correlation coefficient, then construct a regression of this feature on the remaining variables, and use the coefficients of this regression to form a common factor. The results showed that the use of a partial correlation coefficient makes it possible not only to select more significant metrics of the connectivity graph but also to identify hidden connections with the target feature of arithmetic abilities that cannot be detected using paired correlations because of the presence of multicollinearity of features.

In the current reality of big data in area of information industry, issues related to distributed and parallel algorithms for data analysis are relevant. In particular, big data appear at the training stage of machine learning algorithms. Boosting algorithms are a vivid representative of this group of algorithms. When data is large, distributed and parallel processing by boosting type algorithms is possible, but it is associated with sharp increase pf processing time. On the one hand, time increases due to the requirement to ensure the necessary accuracy of problem solutions, on the other, due to the need to transport distributed data and synchronize the entire computation process. Discussions and experiments conducted on this topic are provided.

End-to-end network has become increasingly important in multitasking. One prominent example of this is the growing significance of a driving perception system in autonomous driving. This paper systematically studies an end-to-end perception network for multitasking and proposes several key optimizations to improve accuracy. First, the paper proposes efficient segmentation head and box/class prediction networks based on weighted bidirectional feature network. Second, the paper proposes automatically customized anchor for each level in the weighted bidirectional feature network. Third, the paper proposes an efficient training loss function and training strategy to balance and optimize network. Based on these optimizations, we have developed an end-to-end perception network to perform multitasking, including traffic object detection, drivable area segmentation and lane detection simultaneously, called HybridNets, which achieves better accuracy than prior art. In particular, HybridNets achieves 77.3 mean average precision on Berkeley DeepDrive Dataset, outperforms lane detection with 31.6 mean intersection over union with 12.83 million parameters and 15.6 billion floating-point operations. In addition, it can perform visual perception tasks in real-time and thus is a practical and accurate solution to the multitasking problem. Code is available at https://github.com/datvuthanh/HybridNets .

Compared to contact detection techniques, pavement crack identification with visual images via deep learning algorithms has the advantages of not being limited by the material of detected object, fast speed and low cost. The pavement crack identification performance of typical models or algorithms of transfer learning, encoder-decoder, and generative adversarial networks, were evaluated and compared on public data sets SDNET2018 and Crack Forest Dataset. Transfer learning mainly takes advantage of fine-tuning the architecture-optimized backbones pretrained on large-scale data sets (e.g., ImageNet) to achieve good classification accuracy. Encoder-decoder based algorithms enlarge the contextual receptive field by dilated convolution, taking into account the fact that crack edges, patterns or texture features contribute differently to the identification, which is ignored by transfer learning. Both transfer learning and encoder-decoder rely on accurate crack ground truth annotation. Generative adversarial networks are compatible with other neural network architectures, thus can integrate various frameworks (e.g., transfer learning, encoder-decoder), and algorithms, but the training time is much longer. In patch classification, the fine-tuned transfer-learning models can be equivalent to or even slightly better than the encoder-decoder based algorithms, and the predicting time is faster; In accurate crack location, both encoder-decoder based and generative-adversarial-network based algorithms can achieve pixel-level segmentation. It is expected to realize end-to-end and real-time automatic crack identification on a low computational power platform. Furthermore, a weakly supervised learning framework is proposed, combining transfer-learning and semisupervised generative-adversarial-network. It only needs approximately 10–20% labeled samples of the total to achieve comparable crack classification performance to or even outperform supervised learning algorithms, via fine-tuned backbones and utilizing extra unlabeled samples.