Tom 33 (2023): Zeszyt 1 (March 2023)
Image Analysis, Classification and Protection (Special section, pp. 7-70), Marcin Niemiec, Andrzej Dziech and Jakob Wassermann (Eds.)

Counting and detecting occluded faces in a crowd is a challenging task in computer vision. In this paper, we propose a new approach to face detection-based crowd estimation under significant occlusion and head posture variations. Most state-of-the-art face detectors cannot detect excessively occluded faces. To address the problem, an improved approach to training various detectors is described. To obtain a reasonable evaluation of our solution, we trained and tested the model on our substantially occluded data set. The dataset contains images with up to 90 degrees out-of-plane rotation and faces with 25%, 50%, and 75% occlusion levels. In this study, we trained the proposed model on 48,000 images obtained from our dataset consisting of 19 crowd scenes. To evaluate the model, we used 109 images with face counts ranging from 21 to 905 and with an average of 145 individuals per image. Detecting faces in crowded scenes with the underlying challenges cannot be addressed using a single face detection method. Therefore, a robust method for counting visible faces in a crowd is proposed by combining different traditional machine learning and convolutional neural network algorithms. Utilizing a network based on the VGGNet architecture, the proposed algorithm outperforms various state-of-the-art algorithms in detecting faces ‘in-the-wild’. In addition, the performance of the proposed approach is evaluated on publicly available datasets containing in-plane/out-of-plane rotation images as well as images with various lighting changes. The proposed approach achieved similar or higher accuracy.

Face recognition (FR) is one of the most active research areas in the field of computer vision. Convolutional neural networks (CNNs) have been extensively used in this field due to their good efficiency. Thus, it is important to find the best CNN parameters for its best performance. Hyperparameter optimization is one of the various techniques for increasing the performance of CNN models. Since manual tuning of hyperparameters is a tedious and time-consuming task, population based metaheuristic techniques can be used for the automatic hyperparameter optimization of CNNs. Automatic tuning of parameters reduces manual efforts and improves the efficiency of the CNN model. In the proposed work, genetic algorithm (GA) based hyperparameter optimization of CNNs is applied for face recognition. GAs are used for the optimization of various hyperparameters like filter size as well as the number of filters and of hidden layers. For analysis, a benchmark dataset for FR with ninety subjects is used. The experimental results indicate that the proposed GA-CNN model generates an improved model accuracy in comparison with existing CNN models. In each iteration, the GA minimizes the objective function by selecting the best combination set of CNN hyperparameters. An improved accuracy of 94.5 % is obtained for FR.

Small target detection under a complex background has always been a hot and difficult problem in the field of image processing. Due to the factors such as a complex background and a low signal-to-noise ratio, the existing methods cannot robustly detect targets submerged in strong clutter and noise. In this paper, a local gradient contrast method (LGCM) is proposed. Firstly, the optimal scale for each pixel is obtained by calculating a multiscale salient map. Then, a subblockbased local gradient measure is designed; it can suppress strong clutter interference and pixel-sized noise simultaneously. Thirdly, the subblock-based local gradient measure and the salient map are utilized to construct the LGCM. Finally, an adaptive threshold is employed to extract the final detection result. Experimental results on six datasets demonstrate that the proposed method can discard clutters and yield superior results compared with state-of-the-art methods.

Data compression combined with effective encryption is a common requirement of data storage and transmission. Low cost of these operations is often a high priority in order to increase transmission speed and reduce power usage. This requirement is crucial for battery-powered devices with limited resources, such as autonomous remote sensors or implants. Well-known and popular encryption techniques are frequently too expensive. This problem is on the increase as machine-to-machine communication and the Internet of Things are becoming a reality. Therefore, there is growing demand for finding trade-offs between security, cost and performance in lightweight cryptography. This article discusses asymmetric numeral systems— an innovative approach to entropy coding which can be used for compression with encryption. It provides a compression ratio comparable with arithmetic coding at a similar speed as Huffman coding; hence, this coding is starting to replace them in new compressors. Additionally, by perturbing its coding tables, the asymmetric numeral system makes it possible to simultaneously encrypt the encoded message at nearly no additional cost. The article introduces this approach and analyzes its security level. The basic application is reducing the number of rounds of some cipher used on ANS-compressed data, or completely removing an additional encryption layer when reaching a satisfactory protection level.

The most commonly used public key cryptographic algorithms are based on the difficulty in solving mathematical problems such as the integer factorization problem (IFP), the discrete logarithm problem (DLP) and the elliptic curve discrete logarithm problem (ECDLP). In practice, one of the most often used cryptographic algorithms continues to be the RSA. The security of RSA is based on IFP and DLP. To achieve good data security for RSA-protected encryption, it is important to follow strict rules related to key generation domains. It is essential to use sufficiently large lengths of the key, reliable generation of prime numbers and others. In this paper the importance of the arithmetic ratio of the prime numbers which create the modular number of the RSA key is presented as a new point of view. The question whether all requirements for key generation rules applied up to now are enough in order to have good levels of cybersecurity for RSA based cryptographic systems is clarified.

This paper deals with the finite-time stabilization problem for a class of uncertain disturbed systems using linear robust control. The proposed algorithm is designed to provide the robustness of a linear feedback control scheme such that system trajectories arrive at a small-size attractive set around an unstable equilibrium in a finite time. To this end, an optimization problem with a linear matrix inequality constraint is presented. This means that the effects of external disturbances, as well as matched and mismatched uncertain dynamics, can be significantly reduced. Finally, the performance of the suggested closed-loop control strategies is shown by the trajectory tracking of an unmanned aerial vehicle flight.

Many interconnected systems in the real world, such as power systems and chemical processes, are often composed of subsystems. A decentralized controller is suitable for an interconnected system because of its more practical and accessible implementation. We use the homotopy method to compute a decentralized controller. Since the centralized controller constitutes the starting point for the method, its existence becomes very important. This paper introduces a non-singular matrix and a design parameter to generate a centralized controller. If the initial centralized controller fails, we can change the value of the design parameter to generate a new centralized controller. A sufficient condition for a decentralized controller is given as a bilinear matrix inequality with three matrix variables: a controller gain matrix and a pair of other matrix variables. Finally, we present numerical examples to validate the proposed decentralized controller design method.

Fractional time-invariant compartmental linear systems are introduced. Controllability and observability of these systems are analyzed. The eigenvalue assignment problem of compartmental linear systems is considered and illustrated with a numerical example.

There are two main approaches to tackle the challenge of finding the best filter or embedded feature selection (FS) algorithm: searching for the one best FS algorithm and creating an ensemble of all available FS algorithms. However, in practice, these two processes usually occur as part of a larger machine learning pipeline and not separately. We posit that, due to the influence of the filter FS on the embedded FS, one should aim to optimize both of them as a single FS pipeline rather than separately. We propose a meta-learning approach that automatically finds the best filter and embedded FS pipeline for a given dataset called FSPL. We demonstrate the performance of FSPL on n = 90 datasets, obtaining 0.496 accuracy for the optimal FS pipeline, revealing an improvement of up to 5.98 percent in the model’s accuracy compared to the second-best meta-learning method.

Depression is one of the primary causes of global mental illnesses and an underlying reason for suicide. The user generated text content available in social media forums offers an opportunity to build automatic and reliable depression detection models. The core objective of this work is to select an optimal set of features that may help in classifying depressive contents posted on social media. To this end, a novel multi-objective feature selection technique (EFS-pBGSK) and machine learning algorithms are employed to train the proposed model. The novel feature selection technique incorporates a binary gaining-sharing knowledge-based optimization algorithm with population reduction (pBGSK) to obtain the optimized features from the original feature space. The extensive feature selector (EFS) is used to filter out the excessive features based on their ranking. Two text depression datasets collected from Twitter and Reddit forums are used for the evaluation of the proposed feature selection model. The experimentation is carried out using naive Bayes (NB) and support vector machine (SVM) classifiers for five different feature subset sizes (10, 50, 100, 300 and 500). The experimental outcome indicates that the proposed model can achieve superior performance scores. The top results are obtained using the SVM classifier for the SDD dataset with 0.962 accuracy, 0.929 F1 score, 0.0809 log-loss and 0.0717 mean absolute error (MAE). As a result, the optimal combination of features selected by the proposed hybrid model significantly improves the performance of the depression detection system.

Transition systems (TSs) and Petri nets (PNs) are important models of computation ubiquitous in formal methods for modeling systems. A crucial problem is how to extract, from a given TS, a PN whose reachability graph is equivalent (with a suitable notion of equivalence) to the original TS. This paper addresses the decomposition of transition systems into synchronizing state machines (SMs), which are a class of Petri nets where each transition has one incoming and one outgoing arc. Furthermore, all reachable markings (non-negative vectors representing the number of tokens for each place) of an SM have only one marked place with only one token. This is a significant case of the general problem of extracting a PN from a TS. The decomposition is based on the theory of regions, and it is shown that a property of regions called excitation-closure is a sufficient condition to guarantee the equivalence between the original TS and a decomposition into SMs. An efficient algorithm is provided which solves the problem by reducing its critical steps to the maximal independent set problem (to compute a minimal set of irredundant SMs) or to satisfiability (to merge the SMs). We report experimental results that show a good trade-off between quality of results vs. computation time.

Forecasting the number of hospitalization patients is important for hospital management. The number of hospitalization patients depends on three types of patients, namely admission patients, discharged patients, and inpatients. However, previous works focused on one type of patients rather than the three types of patients together. In this paper, we propose a multi-task forecasting model to forecast the three types of patients simultaneously. We integrate three neural network modules into a unified model for forecasting. Besides, we extract date features of admission and discharged patient flows to improve forecasting accuracy. The algorithm is trained and evaluated on a real-world data set of a one-year daily observation of patient numbers in a hospital. We compare the performance of our model with eight baselines over two real-word data sets. The experimental results show that our approach outperforms other baseline algorithms significantly.