Vol 29, No 4 (2019)

The known quadratic \(NP\)-hard (in the strong sense) \(M\)-variance problem is considered. It arises in the following typical problem of data analysis: in a set of \(N\) objects determined by their characteristics (features), find a subset of \(M\) elements close to each other. For the one-dimensional case, an accelerated exact algorithm with complexity \(\mathcal{O}(N{\kern 1pt} \log{\kern 1pt} N)\) is proposed.

Image segmentation plays an important role in machine vision, image recognition, and imaging applications. Based on the fuzzy c-means clustering algorithm, a kernel-distance-based intuitionistic fuzzy c-means clustering (KIFCM) algorithm is proposed. First, a fuzzy complement operator is used to generate the membership degree whereby the hesitation degree of intuitionistic fuzzy set is generated; second, a kernel-induced function is used to calculate the distance from each point to the cluster center instead of the Euclidean distance; third, a new objective function that includes the hesitation degree is established, and the optimization of the objective function results in new iterative expressions for the membership degree and the cluster center. The proposed KIFCM algorithm is compared with the fuzzy c-means clustering (FCM) algorithm, the kernel fuzzy c-means clustering (KFCM) algorithm, and the intuitionistic fuzzy c-means clustering (IFCM) algorithm in segmenting five images. The experimental results verify the effectiveness and superiority of our proposed KIFCM algorithm.

The article is the second in a series on the current state and prospects of Descriptive Image Analysis, which is the leading branch of the modern mathematical theory of image analysis. Descriptive image analysis is a logically organized set of descriptive methods and models for analyzing and evaluating information in the form of images and for automating knowledge and data extraction from images necessary for making intelligent decisions about real-world scenes displayed and represented in an analyzed image. Problems on making intelligent decisions based on data analysis require formal representation of the source information, ideally, a mathematical model. Image modeling has a long, but not very productive history. Therefore, in the Descriptive Approach to image analysis and understanding (DA), the primary problem is bringing an image to a form suitable for recognition. The DA interprets the sought representation in the form of a descriptive image model (DIM). Due to the extremely complex informational nature and technical features involved in the digital representation of an image, it is impossible to construct a classical mathematical model of an image as an information object. To overcome this complexity and regularize the problem of bringing an image to a form convenient for recognition, a new mathematical object, a DIM is introduced and used in the DA. Models of recognition objects—images—and definitions of transformations over image models are considered. A formalized concept of descriptive image models is proposed. The results can be used to create a basis for methods of transforming and understanding an image as a mathematical object. The article’s main contribution to developing the mathematical theory of image analysis is understanding of an image as an information object and mathematical object.

The generalized spectral-analytical method as a new approach to the processing of information arrays is stated. Some theoretical foundations of this method and its applications in different experimental data analysis problems are given. The method is based on the adaptive expansion of initial arrays in the functional bases belonging to the classical algebraic systems of polynomials and functions of continuous and discrete arguments (Jacobi, Chebyshev, Lagrange, Laguerre, Kravchuk, Charlier, and other polynomials). This approach combines analytical and digital data-processing procedures, thus providing a basis for the universal combined technology for the processing of information arrays. An appreciable part of this review is devoted to video data analysis and pattern-recognition problems. In addition, some relevant applications of this method in biomedical and bioinformation data analysis, recognition, classification, and diagnosis problems are described.

The paper considers the problem of numerical estimation of the closeness of a topical text to the most rational linguistic variant (i.e. semantic pattern or sense standard) of the description of the knowledge fragment it represents without paraphrasing. This problem is relevant when implementing targeted selection of text information by the maximum of the useful semantic component with respect to the tasks solved by the user. Examples of practical applications may include selection of papers for scientific publishing and design of training courses and educational portals. In the suggested solution, the basis of the estimate of the closeness of the text to the semantic pattern is the splitting of the words of each of its phrases into classes by the TF-IDF metric value relative to texts of a corpus preformed by an expert. Abstracts of scientific papers together with their titles are analyzed. The suggested numerical estimate of closeness to the sense standard makes it possible to rank articles by the significance of the described fragments of knowledge regarding a given subject area and by non-redundancy of the description itself. Here, the semantic images of the texts closest to the semantic pattern specify the words with the highest TF-IDF values, which, when placed next to each other in the linear series of a phrase, are, most probably, semantically related and form key combinations with words whose mentioned metric is close to average. To classify word combinations as key ones, the interpretation of the TF-IDF metric, estimating the number of simultaneous occurrences of all words in the analyzed combination into phrases of the individual document, is introduced.

Analysis of criteria for the solvability/regularity of problems and of the correctness of algorithms is applied here to the problem of prediction of the values of numerical variables. It is shown that partial regularity is a necessary and sufficient condition for the solvability of the corresponding system of the classification problems. Cross-validation experiments conducted on several datasets from the field of biomedicine (non-invasive diagnostics of magnesium concentration in blood plasma), bioinformatics (prediction of the protein secondary structure), and solid-state physics (prediction of the properties of high-temperature superconductors) have demonstrated the effectiveness of the developed methods for generating “synthetic” informative numerical features and for increasing the accuracy of prediction of the numerical target variables.

The human pose estimation is one of the most popular research fields. Its current accuracy is satisfactory in some cases, however, there exists a challenge for practical application due to the limited memory and computational efficiency in FPGAs and other hardware. We propose a lightweight module based on the pyramid residual module in this work. We change the convolution mode by using the depth-wise separable convolutions structure. Meanwhile, the channel split module and channel shuffle module are added to change the feature graph dimension. As a result, the parameters of the network are reduced effectively. We test the network on standard benchmarks MPII dataset, our method reduces about 50% of the training storage space while maintaining comparable accuracy. The complexity is simplified from 9 GFLOPs to 3 GFLOPs.

Iris recognition is recognized as one of the most reliable and efficient technique for human identification in the biometric fields. The Gabor filter and local binary pattern (LBP) are widely adopted for feature extraction in face recognition. However, it is difficult to achieve high recognition accuracy when the Gabor filter or LBP is directly applied to iris texture representation. This paper presents an effective iris feature descriptor, which first uses 2D-Gabor filter to extract multi-orientation imaginary (MOI) feature, and then applies uniform LBP for region feature encoding. Thus, the MOI feature-by-point energy is converted into that of the uniform LBP histogram-by-block, during which the distributions of the intra- and inter-class are greatly widened. Such process largely improves distinguishability of MOI features. Finally, the Bhattacharyya distance is adopted for matching. Experimental results on CASIA and JLU iris image databases show that this method performs better for combining MOI features and LBP encoding as compared to their individual function.

In this paper, we present a new deep learning-based human activity recognition technique. First, we track and extract human body from each frame of the video stream. Next, we abstract human silhouettes and use them to create binary space-time maps (BSTMs) which summarize human activity within a defined time interval. Finally, we use convolutional neural network (CNN) to extract features from BSTMs and classify the activities. To evaluate our approach, we carried out several tests using three public datasets: Weizmann, Keck Gesture and KTH Database. Experimental results show that our technique outperforms conventional state-of-the-art methods in term of recognition accuracy and provides comparable performance against recent deep learning techniques. It’s simple to implement, requires less computing power, and can be used for multi-subject activity recognition.

This article presents a new approach to solving the problem of automated detection of myocardial infarction of various localization by electrocardiogram data entries. Only the second standard lead is used in the analysis. The signal in this lead undergoes digital filtering in order to remove low-frequency and high-frequency interference. Then, individual cardio complexes P-QRS-T are extracted from the signal, and the following parameters are calculated for them: minimum value, maximum value, interquartile range, mean absolute deviation, root mean square, mode, and entropy. Using the calculated parameters, a standardized training (learning) dataset is formed. The classifier model represents the k-nearest neighbors algorithm with the Manhattan metric of the distance between the objects and number of neighbors k = 9. After learning, the classifier shows the results by precision pre = 98.60%, by recall rec = 97.34%, by specificity spec = 95.93%, and by accuracy acc = 97.03%. According to the analysis of the obtained results, the suggested classifier model offers certain advantages as compared to existing alternatives.

The paper discusses methodological techniques for increasing the diagnostic value of routine angiographic examinations of patients. It presents algorithms for digital processing of heart video images, which make it possible to quantitatively characterize hemodynamics in the coronary bed by determining the velocity of spread of a contrast agent through the arteries. The proposed approach includes several stages and procedures and takes account of the errors caused by the movement of the arteries due to the mechanical activity of the heart. The paper presents the results of estimating coronary blood-flow velocity in a patient with coronary heart disease, which are compared with computer simulation data. The sources of errors, ways to minimize them, and prospects for using the proposed methodology for effective angiographic diagnosis are discussed.