Accurate prediction of solar wind density fluctuations is essential for space weather forecasting due to their significant impact on satellite operations, power grids, and communication systems. In this work, we present a novel, physics-free forecasting framework that relies exclusively on binary masks of solar surface active regions and historical solar wind density measurements at L1. Our architecture integrates a Graph Neural Network (GNN) to encode the topological structure of binary active region maps (derived from the SDO dataset provided by NASA) with the time series prediction power of a Long Short-Term Memory (LSTM) network for modeling the electron and proton densities at L1 (extracted from the OMNI dataset provided by NASA) from 2012 to 2014. Based on the graph representation of binary masks only, our model simplifies and lowers the number of parameters by a significant degree compared to conventional convolutional approaches, also surpassing their predictive power. Our model demonstrated superior performance over two CNN-based baselines (ConvLSTM and CNN-LSTM). The strength of our solution lies in the use of graph representations to preserve the spatial topology information that pixel-based methods tend to overlook. The results indicate that light topology-preserving models capable of delivering reliable solar wind density predictions are feasible, making it possible to have efficient onboard space weather warning systems.

In this paper, we address the issue of digital camera identification, a key area in digital forensics. This problem is well-known in the literature, with many algorithms proposed for identifying digital cameras based on their fingerprints. However, there is a notable lack of methods that provide fast and accurate digital camera identification, especially given the large image sizes produced by modern digital cameras. Additionally, fingerprints are typically represented as matrices of a size corresponding to the camera’s input images, which can pose storage challenges for forensic centers. Therefore, we propose a method that utilizes a fully convolutional network (FCN) for digital camera identification which is faster than traditional convolutional neural networks (CNN), equipped with fully connected layers. Moreover, we also show that the proposed network can detect lens aberrations, including lens vignetting and distortion. Extensive experimental evaluation conducted on many cameras and images demonstrates the reliability of the proposed method.

We present an innovative method for constructing ensembles of modular networks designed for data classification, based on novel one-class classifier-type structures equipped with a so-called glial driver. This concept is inspired by recent neurobiological discoveries highlighting the significant impact of glial cells on cognitive processes in the human brain. The proposed solution was implemented and validated under real-world conditions within a Polish government ministry to automate document routing. The process of building one-class classifier models is also novel, as it involves training a single structure with the entire training sequence before converting the trained structure into a one-class classifier model using glial cells. This approach demonstrates a substantial reduction in the training time of the modular system and a significant improvement in its performance.

The smart city system, which is a type of enterprise management system (EMS), automatically manages cities and schedules resources efficiently based on spatial data generated by devices, such as the Internet of Things and mobile. However, with the increasing deployment of technologies, including sensor and location-based services, their ever-growing spatial data are no longer managed efficiently by traditional EMS. To overcome this issue, we present SeFrame, which is aspatiallyenabledframework for improving the efficiency of smart city EMS based on a distributed architecture. The framework supports a set of spatial queries, including: The range query, k-nearest neighbors query, and spatial join query. It benefits greatly from using the buffer-enabled partition method to eliminate duplicate results. In each partition, the local index based on combination of the quad-tree and grid index (CQG) significantly improves the spatial query efficiency in memory. CQG manages complex spatial objects, including a point, polygon, and polyline. By taking full advantage of the local index, SeFrame accesses skewed spatial data in constant time. In experiments, we demonstrated that the proposed method delivered superior performance in terms of scalability and query efficiency, in most cases.

With the advance of artificial intelligence (AI), the Internet of Things (IoT), and 5G communication technologies, various kinds of traffic data from diverse devices can be acquired nowadays, and they can help us look into intelligent transportation systems (ITSs) with a new eye. Graph-based machine learning holds out the potential as a powerful tool for modeling complex structural data relationships and also mining both useful information and temporal patterns which could be used for building powerful analytics for ITS construction. Considering the benefit of graph-based machine learning for ITS, some graph-based machine learning methods/architectures have been proposed. Even though these methods have achieved certain success, there exist various scientific and engineering challenges.

The paper proposes a method for classifying and fast retrieving full-disk images of the Sun chromosphere and corona collected by the Solar Dynamics Observatory spacecraft. The method uses a convolutional autoencoder to encode the solar images in the form of a concise semantic hash. The speed of the retrieval comes from the boosting meta-learning to construct a set of fuzzy rules describing the hash parameters. The rules constitute a set of weak classifiers voting for the final image class. This allows for fast retrieving similar images in vast collections of solar images.

The Sun's activity has an enormous impact on the climate and technology on the Earth. The NASA Solar Dynamic Observatory (SDO) spacecraft produces an enormous number of images a day. We propose a technique for fast retrieving and indexing these images. We use image preprocessing, novel edge detection based on a nonparametric Parzen kernel estimator for discovering active regions and a convolutional autoencoder to generate a concise solar image description. Such real-valued hashes can be used to retrieve similar Sun images. Moreover, they can be a basis to classify or predict Solar activity. The experiments showed excellent accuracy and relatively good speed of the proposed method.

Machine learning techniques have achieved promising results in detecting attacks in computer networks, particularly ensemble learning methods, improving individual classifier’s performance. This work focuses on building an ensemble of classifiers to minimize the computational cost to some extent. A diversity-driven pruning method was applied to create stackings using a combination of k-Nearest Neighbors, Decision Trees, Support Vector Machines, and Neural Networks, and validated on six differents datasets. An average accuracy of 99.94% and a reduction in the processing time of 97.34% are reported with heterogeneous ensembles, highlighting the robustness of the proposed approach.

Abstract High resolution remote sensing image cloud removal can bring a lot of convenience for human activities. However, the existing cloud removal algorithms have a variety of disadvantages. First of all, they have the disadvantages of long computing time and large consumption of computing resources. Secondly, the effect of recovery needs to be improved. In order to improve the above two points, a near real‐time effective algorithm is proposed, namely MSAR‐Defognet (multiple scale attention residual network using for cloud remove), which consumes less computing power and space and has superior cloud removal effect. On the one hand, several different large‐scale filters are chosen to extract the weak information effectively, while can save the computing power and shorten the image processing time. On the other hand, the fine‐grained convolution residual block with channel attention mechanism is used to enhance the network's ability to extract cloud features. In addition, a data set which is closer to the real cloud shape and has higher richness to train the cloud removal network, so that the parameters obtained by training have stronger robustness and can adaptively remove clouds with different thickness. Experiments show that, compared with other advanced network models, the network not only has the advantage of fast processing speed, but also has better image restoration effect in high‐resolution remote sensing image restoration. It can meet the requirements of many hard real‐time tasks, so that remote sensing images can play a greater value for human activities.

The Solar Dynamics Observatory delivers data concerning various aspects of the Sun activity. Its Atmospheric Imaging Assembly, performs continuous full-disk observations of the solar chromosphere and corona in seven extreme ultraviolet channels with the 12-second cadence of high-resolution, over 16-megapixel images. In the paper, we create a fast, concise hash to retrieve similar solar images in this vast collection. We use a fully convolutional autoencoder along with a hand-crafted intermediate descriptor.

Point cloud classification is highly dependent on how points' features are extracted and aggregated. The graph-based feature extraction strategies are currently used. Not only point coordinates are taken into consideration but also neighbourhood pair-wise geometrical relations. A newly proposed Convolutional Vicinity Aggregation (CVA) module extends reference solutions with mutual geometrical point relations. Simultaneous convolution of points geometrical interrelations allows a network to retrieve salient features under the permutation-invariance constraint. The resulting hierarchical CVA-based architecture outperforms the state-of-the-art point cloud classification methods on the well-established ModelNet40 dataset. Additional analysis of the CVA module hyper-parameters was also provided in order to support its effectiveness.

The paper proposes a method for classifying and fast retrieving images which uses boosting metalearning to search for the most salient image features. We use local image keypoints as image features. We construct by boosting a set fuzzy rules describing image feature parameters. The rules constitute a set of weak classifiers voting for the final image class. The method can use various image features, engineered and learned by deep learning methods. We checked the methods on some real-world images.

The Solar Dynamics Observatory provides data to research the connected Sun-Earth system and the impact of the Sun on living on the Earth. Its part, the Atmospheric Imaging Assembly, performs continuous full-disk observations of the solar chromosphere and corona in seven extreme ultraviolet channels with the 12-second cadence of high-resolution, over 16-megapixel images. In the paper, we create a fast binary hash to retrieve similar solar images in this vast collection. We use a fully convolutional autoencoder working on preprocessed solar full-disk projections.

Human pose estimation in a gait sequence is an essential step for solving human identification problems, medical diagnosis, monitoring, and rehabilitation. In this paper, a low-cost Kinect V2.0 sensor is used for investigating motion signatures obtained from normal healthy adults. The purpose of this study is to determine the accuracy and reliability of observational assessments of spatio-temporal features. A novel approach for human detection and tracking is proposed, which involves gait feature learning principles from depth and RGB video. In the first step, a human object from the depth image is extracted using the proposed semi-dynamic object tracking algorithm, and a stick model is generated using body aspect ratios to extract hip angles. In the second step, the gait energy image (GEI) representation is utilized for training a 2D Convolutional Neural Network (AlexNet) for automatic feature extraction. A key point detection algorithm is proposed for estimating knee, hip, and ankle joints from RGB gait videos. The reliability analysis of motion signatures is performed using various statistical methods to ensure feature learning for multi-scenario applications. The statistical results are promising for evaluating the methods which influence the inter-record differences among motion signatures.

Image compression is one of the essential methods of image processing. Its most prominent advantage is the significant reduction of image size allowing for more efficient storage and transfer. However, lossy compression is associated with the loss of some image details in favor of reducing its size. In compressed images, the deficiencies are manifested by noticeable defects in the form of artifacts; the most common are block artifacts, ringing effect, or blur. In this article, we propose three models of fully convolutional networks with different configurations and examine their abilities in reducing compression artifacts. In the experiments, we research the extent to which the results are improved for models that will process the image in a similar way to the compression algorithm, and whether the initialization with predefined filters would allow for better image reconstruction than developed solely during learning.

Sequential pattern mining can discover many interesting phenomena such as bank transactions, web page requesting sequences, customer behavior, etc. There have been many frequent itemset mining algorithms proposed so far, yet it is still a challenging task. In this paper, we propose a deep learning architecture for discovering closed sequences. The U-Net network is trained with random, synthetic sequences and, afterward, is able to discover unknown (not seen during training) sequences. The proposed solution is faster than traditional sequential data mining methods for longer sequences.

A new speckle suppression algorithm is proposed for high-resolution synthetic aperture radar (SAR) images. It is based on the nonlocal means (NLM) filter and the modified Aubert and Aujol (AA) model. This method takes the nonlocal Dirichlet function as a linear regularization item, which constructs the weight by measuring the similarity of images. Then, a new despeckling model is introduced by combining the regularization item and the data item of the AA model, and an iterative algorithm is proposed to solve the new model. The experiments show that, compared with the AA model, the proposed model has more effective performance in suppressing speckle; namely, ENL and DCV measures are 21.75% and 4.5% higher, respectively, than for NLM. Moreover, it also has better performance in keeping the edge information.

Abstract Image vectorisation is a fundamental method in graphic design and is one of the tools allowing to transfer artist work into computer graphics. The existing methods are based mainly on segmentation, or they analyse every image pixel; thus, they are relatively slow. We introduce a novel method for fast line drawing image vectorisation, based on a multi‐scale second derivative detector accelerated by the summed‐area table and an auxiliary grid. Image is scanned initially along the grid lines, and nodes are added to improve accuracy. Applying inertia in the line tracing allows for better junction mapping in a single pass. Our method is dedicated to grey‐scale sketches and line drawings. It works efficiently regardless of the thickness of the line or its shading. Experiments show it is more than two orders of magnitude faster than the existing methods, not sacrificing accuracy.

Automatic detection of abnormal behaviour of computer network users is a desirable and hard to achieve feature.We show that convolutional neural networks can classify users in local computer networks based on features of web pages which were requested by a user (e.g.URL address, URL category, the day of week or time when the web page was visited).We demonstrate our approach on data collected from a firewall over an eight-month period.This network traffic meta-data allowed to achieve satisfactory classification accuracy on unseen, future network traffic data.

Big data term refers to different variations of large datasets to complex to be processed by traditional computing methods. The paper presents a system for retrieving images in relational databases in a distributed environment. Content of the query image and images in the database is compared using global color information and local image keypoints. Image keypoints are indexed by fuzzy sets directly in a relational database. To distribute the process to several machines we use the Apache Hadoop software framework with HDFS.

In this paper we present a new method for content-based searching large image databases by comparing content of a query image and images stored in a database. The algorithm consists of three main steps: feature extraction, indexing and system learning. The feature extraction stage is based on two types of features (SURF keypoints and color). For indexing we use the k-means algorithm and for system learning we applied differential evolution. This last step is very important, and significantly improves the results. The presented algorithm can be easily modified, by changing its components (feature extractor or clustering algorithm).

This paper describes a novel method of image key-point descriptor indexing and comparison used to speed up the process of content-based image retrieval as the main advantage of the dictionary-based representation is faster comparison of image descriptors sets in contrast to the standard list representation. The proposed method of descriptor representation allows to avoid initial learning process, and can be adjusted taking into consideration new examples. The presented method sorts and groups components of descriptors in the process of the dictionary creation. The ordered structure of the descriptors dictionary is well suited for quick comparison of images by comparing their dictionaries of descriptors or by comparing individual descriptors with the dictionary. This allows to skip a large part of operations during descriptors comparison between two images. In contrast to the standard dictionary, our method takes into account the standard deviation between the image descriptors. This is due to the fact that almost all descriptors generated for the points indicating the same areas of the image have different descriptors. Estimation of the similarity is based on the determined value of the standard deviation between descriptors. We assume that proposed method can speed up the process of descriptor comparison. It can be used with many solutions which require high-speed operations on the image e.g. robotics, or in software which computes panoramic photography from scrap images and in many others.

Most methods constituting the soft computing concept can not handle data with missing or unknown feature values. Neural networks are able to perfectly fit to data and fuzzy logic systems use interpretable knowledge. In the paper we incorporate rough set theory to neuro-fuzzy system of very specific type. This results in learning systems which can work when the set number of available feature values is changing. To achieve better accuracy learning systems can be combined into larger ensembles. In the paper the AdaBoost metalearning is used to create an ensemble of learning systems. The rough-neuro-fuzzy systems use knowledge comprised in the form of fuzzy rules to perform classification. Simulations on a well-known benchmark give legitimacy to use the method in real world applications.

A method frequently used in classification systems for improving classification accuracy is to combine outputs of several classifiers. Among various types of classifiers, fuzzy ones are tempting because of using intelligible fuzzy if-then rules. In the paper we build an AdaBoost ensemble of relational neuro-fuzzy classifiers. Relational fuzzy systems bond input and output fuzzy linguistic values by a binary relation; thus, fuzzy rules have additional, comparing to traditional fuzzy systems, weights - elements of a fuzzy relation matrix. Thanks to this the system is better adjustable to data during learning. In the paper an ensemble of relational fuzzy systems is proposed. The problem is that such an ensemble contains separate rule bases which cannot be directly merged. As systems are separate, we cannot treat fuzzy rules coming from different systems as rules from the same (single) system. In the paper, the problem is addressed by a novel design of fuzzy systems constituting the ensemble, resulting in normalization of individual rule bases during learning. The method described in the paper is tested on several known benchmarks and compared with other machine learning solutions from the literature.

Most methods constituting the soft computing concept can not handle data with missing or unknown features. Neural networks are able to perfectly fit to data and fuzzy logic systems use interpretable knowledge. To achieve better accuracy learning systems can be combined into larger ensembles. In this paper we combine logical neuro-fuzzy systems into the AdaBoost ensemble and extract fuzzy rules from the ensemble. The rules are used in rough-neuro-fuzzy classifier which can operate on data with missing values. The rough systems perform very well on these rules which was illustrated on a well known benchmark. The features were being removed to check the performance on incomplete data sets.

Takagi-Sugeno fuzzy systems are very common learning systems. The paper is about building classification ensembles from them and merging resulting rule bases. When merged, the rule base is more intelligible and easier to process. The merging is possible thanks to a modification of TS systems. Numerical simulations show that the modified systems perform very well

Classification accuracy is nearly always improved after combining many systems. One of the most popular methods of multiple classification is boosting. In the paper we develop a method for merging fuzzy rule bases of neuro-fuzzy systems constituting an ensemble trained by the boosting algorithm.

Boosting is a method for learning combined classifiers. In a boosting ensemble of classifiers trained by the backpropagation algorithm, the learning rate takes much smaller value comparing with the backpropagation applied alone. We propose a method which overcomes the above drawback and test it on neuro-fuzzy systems constituting a classifier ensemble using some well known benchmarks.

In this paper, we present a neuro-fuzzy structure of the hierarchical prioritized structure (HPS) proposed by Yager. The HPS allows for easy hierarchization of a fuzzy rule-base. Our neuro-fuzzy system can be learned by the backpropagation algorithm and is relatively computationally efficient.