Vol. 39 · No. 3

The recent attempts to use supervised learning techniques for process model matching have yielded below par performance. To address this issue, we have transformed the well-known benchmark correspondences to a readily usable format for supervised learning. Furthermore, we have performed several experiments using eight supervised learning techniques to establish that imbalance in the datasets is the key reason for the abysmal performance. Finally, we have used four data balancing techniques to generate balanced training dataset and verify our solution by repeating the experiments for the four datasets, including the three benchmark datasets. The results show that the proposed approach increases the matching performance significantly.

Perceptual quality assessment is a tough task especially in the absence of reference information. No-reference image quality assessment is more challenging than full-reference or reduced reference methods, as the system has to model the different image distortions in the form of a quality score. Most of the approaches are based on handcrafted features which are based on natural scene statistics and are specific to some distortion types. These approaches provide high correlation with human opinion score for datasets containing specific distortions, but they fail to generalize well in scenarios were multiple distortions or real-time distortions are present in images. Deep learning algorithms, on the other hand, demonstrated their abilities to learn expert features with better discriminatory power for various classification and regression tasks. It is a big challenge to use those deep learning methods for image quality assessment as the image datasets with human opinion score are very small and cannot be used effectively to train a deep learning algorithm. We experimented with activations of different deep layers of thirteen pre-trained models and checked for their suitability for the task of no-reference quality assessment. Fine-tuning of these models on quality assessment datasets provided even better performance. A Gaussian process regression model is trained on these activations to perform the quality assessment and it provided state-of-the-art performance. Cross-dataset validation demonstrated its performance further and also provided further prospects of research in this direction.

Using service function chains, Internet Service Providers can customize the use of service functions that process the network flows belonging to their customers. Each network flow is injected into a service chain according to the flow features. Since most of the malicious applications try not to get the proper analysis by imitating some valid and famous applications, classification based on simple flow features may waste processing power by using inappropriate service chains for evasive flows. In this paper, we have explored an application-aware classification approach using machine learning methods. Using CatBoost as a machine learning method, a model is created and used for traffic classification. We have provided some statistical reports on how this approach is compared with simple flow feature-based approaches in malicious environments and how feature selection can impact classification correctness. Choosing the most suitable number of features at the right time can beat traditional approaches in classification quality and provide better results in the service function chaining environment.

Traffic light recognition in urban environments is crucial for vehicle control. Many studies have been devoted to recognizing traffic lights. However, existing recognition methods still face many challenges in terms of accuracy, runtime and size. This paper presents a novel robust traffic light recognition approach that takes into account these three aspects based on image processing and deep learning. The proposed approach adopts a two-stage architecture, first performing detection and then classification. In the detection, the perspective relationship and the fractal dimension are both considered to dramatically reduce the number of invalid candidate boxes, i.e. region proposals. In the classification, the candidate boxes are classified by SqueezeNet. Finally, the recognized traffic light boxes are reshaped by postprocessing. Compared with several reference models, this approach is significantly competitive in terms of accuracy and runtime. We show that our approach is lightweight, easy to implement, and applicable to smart terminals, mobile devices or embedded devices in practice.

The problem of measuring similarity between sentences is crucial for many applications in Natural Language Processing (NLP). Most of the proposed approaches depend on similarity of words in sentences. This research considers semantic relations between words in calculating sentence similarity. This paper uses Discourse Representation Structure (DRS) of natural language sentences to measure similarity. DRS captures the structure and semantic information of sentences. Moreover, the estimation of similarity between sentences depends on semantic coverage of relations of the first sentence in the other sentence. Experiments show that exploiting structural information achieves better results than traditional word-to-word approaches. Moreover, the proposed method outperforms similar approaches on a standard benchmark dataset.

Differential evolution (DE) is an efficient population-based search algorithm with good robustness, however, it is challenged to deal with high-dimensional problems. In this paper, we propose an improved multi-population differential evolution with best-and-current mutation strategy (mDE-bcM). The population is divided into three subpopulations based on the fitness values, each of subpopulations uses different mutation strategy. After crossover, mutation and selection, all subpopulations are updated based on the new fitness values of their individuals. An improved mutation strategy is proposed, which uses a new approach to generate base vector that is composed of the best individual and current individual. The performance of mDE-bcM is evaluated on a set of 19 large-scale continuous optimization problems, a comparative study is carried out with other state-of-the-art optimization techniques. The results show that mDE-bcM has a competitive performance compared to the contestant algorithms and better efficiency for large-scale optimization problems.

The paper investigates parallel data processing in a hybrid CPU+GPU(s) system using multiple CUDA streams for overlapping communication and computations. This is crucial for efficient processing of data, in particular incoming data stream processing that would naturally be forwarded using multiple CUDA streams to GPUs. Performance is evaluated for various compute time to host-device communication time ratios, numbers of CUDA streams, for various numbers of threads managing computations on GPUs. Tests also reveal benefits of using CUDA MPS for overlapping communication and computations when using multiple processes. Furthermore, using standard memory allocation on a GPU and Unified Memory versions are compared, the latter including programmer added prefetching. Performance of a hybrid CPU+GPU version as well as scaling across multiple GPUs are demonstrated showing good speed-ups of the approach. Finally, the performance per power consumption of selected configurations are presented for various numbers of streams and various relative performances of GPUs and CPUs.

Data exposure and privacy violations may happen when data is exchanged between organizations. Data anonymization gives promising results for limiting such dangers. In order to maintain privacy, different methods of k-anonymization and l-diversity have been widely used. But for larger datasets, the results are not very promising. The main problem with existing anonymization algorithms is high information loss and high running time. To overcome this problem, this paper proposes new models, namely Improved k-Anonymization (IKA) and Improved l-Diversity (ILD). IKA model takes large k-value using a symmetric as well as an asymmetric anonymizing algorithm. Then IKA is further categorized into Improved Symmetric k-Anonymization (ISKA) and Improved Asymmetric k-Anonymization (IAKA). After anonymizing data using IKA, ILD model is used to increase privacy. ILD will make the data more diverse and thereby increasing privacy. This paper presents the implementation of the proposed IKA and ILD model using real-time big candidate election dataset, which is acquired from the Madhya Pradesh State Election Commission, India (MPSEC) along with Apache Storm. This paper also compares the proposed model with existing algorithms, i.e. Fast clustering-based Anonymization for Data Streams (FADS), Fast Anonymization for Data Stream (FAST), Map Reduce Anonymization (MRA) and Scalable k-Anonymization (SKA). The experimental results show that the proposed models IKA and ILD have remarkable improvement of information loss and significantly enhanced the performance in terms of running time over the existing approaches along with maintaining the privacy-utility trade-off.

The main aim of the proposed paper is to design a new software tool for modelling and control of discrete-event and hybrid systems using Arduino and similar microcontrollers. To accomplish these tasks a new tool called PN2ARDUINO based on Petri nets is proposed which is able to communicate with the microcontroller. Communication with the microcontroller is based on the modified Firmata protocol hence the control algorithm can be implemented on all microcontrollers that support this type of protocol. The developed software tool has been successfully verified in control of laboratory systems. It can also be used for education and research purposes as it offers a graphical environment for designing control algorithms for hybrid and mainly discrete-event systems. The proposed tool can improve education and practice in the field of cyber-physical systems (Industry 4.0).

This paper deals with the issue of geometry representation of voxelized three-dimensional scenes using hierarchical data structures. These include pointerless Sparse Voxel Octrees that have no pointers on child nodes and allow a compact binary representation. However, if necessary, there is a possibility to reconstruct these pointers for rapid traversing. Sparse Voxel Directed Acyclic Graphs added 32-bit pointers to child nodes and merging of common subtrees, which can be considered lossless compression. By merging common subtrees, no decompression overhead occurs at the time of traversing. The hierarchical data structure proposed herein – the Pointerless Sparse Voxel Directed Acyclic Graph – incorporates the benefits of both – pointerless Sparse Voxel Octrees (by avoiding storing pointers on child nodes) and Sparse Voxel Directed Acyclic Graphs (by allowing the merging of common subtrees due the introduction of labels and callers). The proposed data structure supports the quick and easy reconstruction of pointers by introducing the Active Child Node Count. It also potentially allows Child Node Mask compression of its nodes. This paper presents the proposed data structure and its binary-level encoding in detail. It compares the effectiveness of the representation of voxelized three-dimensional scenes (originally represented in OBJ format) in the proposed data structure with the data structures mentioned above. It also summarizes statistical data providing a more detailed description of the various parameters of the data structure for different scenes stored in multiple resolutions.