Activiteit

Compared to the YOLOv5 model, the enhanced version achieved frame rates of 75 MB, 969%, 599%, and 385 frames per second, respectively; a corresponding increase of 103%, 713%, and 12647%, respectively, underscores the model’s applicability to the cotton foreign fiber detection system used on an industrial production line.

The manufacturing sector consistently experiences a high rate of printing flaws. Despite existing research into the identification of printing faults, the robustness and applicability of such detection methods remain under-appreciated. The accuracy and practicality of current printing defect detection are jeopardized by external environmental factors, specifically the fluctuations in illuminance and the presence of noise. This research employs scale-adaptive template matching and image alignment to formulate a new approach for detecting printing defects. The research method begins by employing a convolutional neural network (CNN) to extract deep feature vectors from low-resolution versions of the template and target images in an adaptive manner. A novel matching metric, feature map cross-correlation (FMCC), is developed to compare template and target image feature maps. A refined location method subsequently determines the exact matching position. Finally, the image that matches and the template are sent to the image alignment module to ascertain the presence of printing defects. The experimental outcomes highlight the proposed method’s 93.62% accuracy in rapidly and precisely locating defects. Proving its efficacy, our method achieves top-tier defect detection performance while simultaneously excelling in real-time detection and interference resistance.

Deep learning presents a multifaceted challenge in Human Activity Recognition (HAR), and 1D Convolutional Neural Networks (1D CNNs) have become a prevalent solution. Data is processed by these networks to extract activity-classifying features, resulting in high performance. Undeniably, the act of comprehending and describing the learned features of these networks represents a considerable challenge. This paper demonstrates a novel eXplainable Artificial Intelligence (XAI) methodology for generating visual explanations of features learned by one-dimensional convolutional neural networks (1D CNNs) during training, through the use of t-Distributed Stochastic Neighbor Embedding (t-SNE). This technique offers insights into the decision-making process, presenting the information gleaned from the model’s deepest level prior to the classification. The learned characteristics within one dataset are demonstrated to be adaptable for the purpose of distinguishing human activities in different data collections in our results. Our trained networks’ performance was notably strong across two public databases, evidenced by an accuracy of 0.98 on the SHO dataset and 0.93 on the HAPT dataset. This investigation introduces a visualization method that effectively detects bias and clarifies the causes of incorrect predictions. A new type of XAI application, detailed in this work, enhances the robustness and practicality of CNN models in real-world deployments.

The expanding array of linked devices has enabled the development of fresh applications in a multitude of areas. In addition, technologies like cloud and fog computing, essential to these applications, face obstacles in supplying adequate processing resources for various applications due to the high dynamism of the underlying networks and the wide variety of heterogeneous devices involved. Within this article, the existing literature on resource allocation in the fog-cloud continuum is analyzed, examining diverse approaches in conjunction with differing strategies and network characteristics. In addition to discussing resource allocation, we analyze the key drivers, including energy consumption levels, latency issues, financial costs, and network traffic. To finalize, we pinpoint the open research questions and suggest promising future directions. Edge computing and resource allocation researchers and practitioners will find this survey article a valuable point of reference.

Magnetic particle spectroscopy (MPS)-based biosensors have experienced a surge in popularity since their initial 2006 report, with the past decade witnessing their significant development. Among the many applications of MPS are disease diagnosis and the detection of foodborne pathogens. Within this work, a comparative analysis of MPS platforms, focusing on dual-frequency and single-frequency driving field implementations, was conducted. Multi-functional magnetic nanoparticles (MNPs), when used in conjunction with MPS, have been extensively investigated as a versatile platform for the detection of a wide range of biomarkers. Nanoprobes, constituted by surface-functionalized magnetic nanoparticles, selectively bind and label target analytes present in liquid samples. A study investigated the theoretical underpinnings and operational mechanisms of various MPS platforms, allowing for bioassay implementation using either volumetric or surface-based approaches. This review, in its further analysis, points to notable implementations of MPS platforms in biomedical and biological contexts. Recent years have witnessed the emergence of a variety of different point-of-care (POC) medical devices incorporating MPS technology, as independently reported by numerous groups around the world. The anticipated broader implementation of MPS point-of-care devices is fueled by their high detection sensitivity, uncomplicated assay procedures, and economical cost per run. The growth of telemedicine, in conjunction with remote monitoring, has fueled a greater demand for point-of-care devices, enabling patients to receive health assessments and obtain results within their own homes. In concluding this review, we analyze the opportunities and hurdles for POC and MPS devices, given the escalating need for fast, inexpensive, highly sensitive, and user-friendly instruments.

This paper explores the concept of a microwave/millimeter wave aperture-sharing antenna. Grounded vias, arranged orthogonally, comprise the construction of the 35 GHz slot-loaded half-mode substrate-integrated waveguide (HMSIW) antenna. These vias are employed to build two separate sets of 1.4 GHz MMW substrate-integrated dielectric resonator antenna (SIDRA) arrays. To achieve shared-aperture integration of the MW antenna and MMW arrays, a proposed partial structure reuse strategy is implemented. This increases space efficiency and enables dual-polarized beam steering in the MMW band, which is an essential requirement for multiple-input multiple-output (MIMO) applications. To verify the integrated antenna prototype, it was manufactured and its performance was measured. An antenna operating at 35 GHz, with a relative bandwidth of 34% and a frequency range of 344-356 GHz, displays a peak antenna gain of 534 dBi. Dual-polarization and 45-degree beam steering are key features of the 28 GHz antenna arrays. The dual-band antenna, with its extremely compact structure, is employed in 5G mobile communication terminals.

Physically unclonable functions represent a promising solution for the authentication of resource-constrained IoT devices, as they avoid storing secret data in non-volatile memories, only generating a key when the application actively requires it. Undeniably, minimizing the forecastability of physically unclonable functions is crucial. We aim to determine the optimal method for building a physically unclonable function in this work. A field-programmable gate array (FPGA) has been used to develop a ring oscillator physically unclonable function, based on the principle of comparing oscillators in pairs. Oscillator frequencies are demonstrably affected by their position in the FPGA, showing substantial divergence, especially when considering oscillators in distinct slice types. Moreover, the impact of the selected ring oscillator positions on the performance of the physically unclonable function has been investigated, and we present five strategies for optimizing oscillator placement. Two strategies from the proposed set demonstrate a high level of uniqueness, reproducibility, and identifiability, making them ideal choices for authentication applications. In closing, we evaluated the reproducibility of the most effective strategy under varying voltage and temperature conditions, confirming its stability over the analyzed range.

To address continuum mechanical challenges, material models are indispensable. Application-specific test environments are instrumental in establishing the parameters contained within these models. Theoretically constructed models, and subsequently their required parameters, exhibit increasing complexity, for example, by considering higher-order derivatives of motion, such as the strain or strain rate. baf-a1 inhibitor Consequently, experimental data must be employed to derive the strain rate behavior. The prevalent method in solid experimental mechanics for accomplishing this task using image data is digital image correlation. Optical flow methods, adaptable to the underlying motion estimation problem, can also be used as an alternative. To estimate strain rate fields robustly, we introduce an optical flow approach that implements higher-order spatial and trajectorial regularization. Unlike higher-order purely spatial variational techniques, the proposed approach enables more precise calculations of displacement and strain rate fields. In the face of difficult optical conditions, the shear cutting experiment’s experimental data, which showcased complex deformation patterns, ultimately demonstrates the procedure.

Data-driven localization within indoor environments necessitates a large dataset for training the predictive model to ensure high accuracy. However, substantial human input is required for WiFi-based data collection to assemble substantial datasets, as the received signal strength (RSS) measurements are vulnerable to interference from obstacles and other elements. Employing an extendGAN+ approach, this paper proposes upsampling with the Dirichlet distribution to boost location prediction accuracy with limited samples, incorporating a transferred WGAN-GP for synthetic data generation, and implementing a data quality filtering module.