Vol. 40 · No. 6

The work is devoted to the description of the question-answer system QA-RiskPanel, which provides means of determining and forecasting the risks related to computer attacks. The QA-RiskPanel system uses a constantly updated database of previous computer attacks as a source of knowledge. We thus guarantee the most up-to-date risk prediction. The ontological approach to the formalization of the object domain allows the analysis of risks at various levels of specification/generalization. In this paper we provide a model-theoretic formalization of the Knowledge Base of the described object domain. Then we describe the classification of question types, which are probabilistic in this system. Finally we present algorithms for finding the answers to all question types of our classification.

Owing to the favorable performance of deep neural networks for 3D shape recognition, an increasing number of researchers are interested in designing novel 3D shape descriptors. However, the relationship between multiple views and point clouds needs to be further elucidated. We propose a multimodal method that combines the features of point clouds and multiple views, i.e., point-view relation neural network embedded with both attention mechanism and Radon transform, to obtain better descriptors. First, a two-dimensional linear Radon transform is performed to investigate linear and color features in multiple views, and the features are used as the input of our network to enable significant distinctions between different views. Moreover, a convolutional block attention module is adopted to enhance the features of point clouds and hence improve the expression ability of feature descriptors. The effectiveness of the proposed method is verified using ModelNet40 and ModelNet10 datasets. Experimental results show that our method can effectively improve the capability of feature extraction and expression as well as achieve state-of-the-art performance on two well-known 3D datasets.

Building segmentation from aerial images is essential in applications such as facilitating urban planning and estimating the population. Fully convolutional networks (FCNs) and especially UNET have achieved promising results in segmentation problems, after deep learning methods have significantly advanced the performance of many computer vision problems. However, in Convolutional Neural Networks (CNNs) with the standard convolution operations, there are problems such as the overfitting and precise extraction of the boundaries of the objects with different sizes and shapes. In this study, we have used Inception blocks with UNET to enhance feature extraction by implementing two-level Inception approach covering the entire encoding stage. In the proposed architecture, structured form of dropout (DropBlock) is used to prevent overfitting, and spatial/channel attention modules are applied to enhance important features by focusing key areas. We evaluate the proposed INCSA-UNET architecture on publicly available Massachusetts dataset and apply two fold cross-validation experiments for better analyzes. The experimental results show that the proposed architecture does not significantly increase the number of parameters of UNET and has a significant improvement in terms of F1 and Kappa quantitative measures.

During the Covid pandemic, the importance of wearing mask has been noted globally. Additionally, crowded human clusters facilitated the transmission of the virus, which brings up the need for new systems for monitoring such situations. To address such issues, this research proposes an object recognition visual system based on deep learning to monitor the mask-wearing in a certain space and the control of the number of people indoors as an important tool during an epidemic. This research mainly investigates two types of identification. The first is to monitor whether people entering the site wear a mask at the entrance and exit of the field, and the second is to count the number of people entering a specific area. Experimental results show that by utilising the visual sensor, it is possible to detect and identify the people who frequently enter and exit in real-time. An advanced transfer learning approach has been employed to achieve the best discrimination performance. The actual training results prove that the migration learning Mask R-CNN algorithm produced by this method and the original Mask R-CNN algorithm have increased the mAP by 3

The main goal of this paper is to perform an extensive analysis of the accuracy of six selected cloud-based facial emotion recognition services on three facial images datasets. The evaluation was performed on more than 33 000 images depicting eight different emotions. Results show that emotion recognition services show a varying level of accuracy over different types of datasets, having a lower accuracy for images of lower quality, but performing considerably better for images taken in ideal conditions. Based on these results we believe that cloud-based facial emotional recognition services do not have the expected accuracy for some use cases and therefore must be selected with care when developing a system that relies on emotion-based interactions.

Clustering is an important data mining technique that helps to detect hidden structures and patterns in the data. K-means algorithm is one of the most popular and widely used partitional clustering algorithms. It is a simple and efficient method but has several shortcomings. One major drawback of traditional K-means is that it selects initial centroids randomly, resulting in low-quality clusters. Various K-means extensions are designed to solve the issue of the random initial centroid. A novel density-based K-means (DK-means) algorithm is recently proposed that uses density-parameters for selecting initial centroids. It outperforms K-means in terms of accuracy at the cost of time. In this research, we present an efficient density-based K-means algorithm (EDK-means) that uses advance data structures and significantly reduces the DK-means algorithm's execution time. Furthermore, we rigorously evaluated the performance of density-based K-means on different challenging real-world datasets and compared it with traditional K-means. The experimental results are promising and show that density-based K-means outperforms K-means. It converges more rapidly than basic K-means, and it works well for the datasets with different cluster sizes.

OpenAPI has become a dominant standard for documentation in the service-oriented software industry. OpenAPI is used in many analysis and reengineering approaches for RESTful service and microservice-based systems. An OpenAPI document has several components that are usually filled by humans using natural language (e.g. description of a certain functionality). Thus, subjectivity may lead to inconsistencies and ambiguities. Understanding what an API does is a challenging question. As a consequence, this issue could hinder developers from identifying the functionality of APIs, after reading all its components. Along this line, we argue that developers will be provided with supportive tools to find those APIs that better suit their needs. In this paper, we propose a step towards creating these kinds of tools by empirically analyzing a set of 2 000 OpenAPI documents with the goal of extracting the main topics of an API using three topic modeling algorithms. To address this issue, we focus on three tasks: i) determine which component of an OpenAPI document provides the most meaningful information, ii) compare three state-of-the-art topic modeling algorithms, and iii) determine the optimal number of topics to represent an API. Our findings show that the best results could be obtained from the Description component by using the Non-negative Matrix Factorization (NMF) or Latent Semantic Indexing (LSI) algorithms. To help developers find services in the OpenAPI directory, we also propose a prototype tool to explore the OpenAPI documents and analyze extracted topics to assess if the APIs meet developers needs.

We study two problems called the Traveling Repairman Problem (TRP) and Traveling Salesman Problem (TSP). The TRP wants to minimize the total time for all customers that have to wait before being served, while the TSP aims to minimize the total time to visit all customers. In this sense, the TRP takes a customer-oriented view, whereas the TSP is server-oriented. In the literature, there exist numerous algorithms that are developed for two problems. However, these algorithms are designed to solve each problem independently. Recently, Multifactorial Evolutionary Algorithm (MFEA) has been a variant of Evolutionary Algorithm (EA) aiming to solve multiple optimization tasks simultaneously. The MFEA framework has yet to be fully exploited, but the realm has recently attracted much interest from the research community. This paper proposed a new approach using the MFEA framework to solve these two problems simultaneously. The MFEA has two tasks simultaneously: the first is solving the TRP problem, and the second is solving the TSP. Experiment results show the efficiency of the proposed MFEA: 1. for small instances, the algorithm reaches the optimal solutions of both problems; 2. for large instances, our solutions are better than those of the previous MFEA algorithms.

This paper proposes an image enhancement algorithm based on the theory of image segmentation and image frequency. First, a mathematical model corresponding to the pixel frequency is established by using the difference of pigment values between the pixels in the image and the surrounding pixels (i.e., pixel receptive field). Then, the image is divided into the low frequency region (background area), low-medium frequency region (foreground area), medium-high frequency region (target area) and high frequency region (detail area) by a pixel frequency characteristic graph. Gamma correction, MSRCR, MSR, top hat+bottom hat are used for image enhancement for each area, and then the parts are merged. Three indicators of PSNR, SSIM, and MSE are introduced to evaluate the quality of the enhanced image. The results show that the image enhanced by this algorithm has the highest PSNR and SSIM values and the lowest MSE value, indicating that the enhancement effect of this algorithm is better. Compared with traditional algorithms, the image enhancement algorithm in this paper produces higher image quality and richer details.

In this paper, we investigate the popular centrality-based approaches to find a set of critical nodes whose deletion causes the most disconnectivity in the network. Demonstrating the weak points of these approaches which only consider a ranking factor, we propose an Enhanced Critical Node Detection (ECND) method which can work with any kind of ranking score by considering the structure of a network. We have designed a set of experiments using 24 different artificial and real-world networks, varying in the number of vertices and number of edges. Using two different objective functions including the number of connected components and the weighted average size of the connected components, experimental results show outperformance of ECND in comparison to all 8 other methods.