A document by Baptiste Daire. Click on the document to view its contents.

Scene boundary detection is crucial in applications like scene segmentation and video skimming, but the presence of ad clips makes it difficult. There are some existing methods based on visual, audiovisual and audio-only features. The existing audio-only feature-based method depends on short silences between program and ad or between ads. Still, short silences can also be present inside the program, affecting performance. So we proposed a detection method for ad shots using MFCC features. MFCC is a basic speech feature based on short-term spectral. We used the average of MFCC for a shot and determined the threshold for detection purposes. We used four episodes of The Big Bang Theory as our dataset. Our method is the first method that uses MFCC features for ad detection and explores the relation between video contents and MFCC features.

Electromagnetic (EM) body models predict the impact of human presence and motions on the Radio-Frequency (RF) field originated from wireless devices nearby. Despite their accuracy, EM models are time-consuming methods which prevent their adoption in strict real-time computational imaging problems and Bayesian estimation, such as passive localization, RF tomography, and holography. Physics-informed Generative Neural Network (GNN) models have recently attracted a lot of attention thanks to their potential to reproduce a process by incorporating relevant physical laws and constraints. They can be used to simulate or reconstruct missing data or samples, reproducing EM propagation effects, approximated EM fields and learn a physics-informed data distribution, i.e., a Bayesian prior. The paper discusses two popular techniques, namely variational auto-encoders (VAEs) and generative adversarial networks (GANs), and their adaptations to incorporate relevant EM body diffraction concepts. Proposed EM-informed generative models are verified against classical EM tools driven by diffraction theory and validated on real data. Physics-informed generative machine learning represents a multidisciplinary research area weaving together physical/EM modelling, signal processing and artificial intelligence (AI): the paper explores emerging opportunities of GNN tools targeting real-time passive RF sensing in multiple-input multiple-output (MIMO) communication systems. Proposed generative tools are designed, implemented and verified on resource constrained wireless devices being members of a Wireless Local Area Network (WLAN).

In analog computing, multimode fibers (MMFs) have been explored for their potential in performing computations by exploiting the inherent parallelism of light propagation through different transverse modes. However, traditional approaches using MMFs encounter challenges due to limitations in spatial-domain encoding, which limit computational speed and accuracy. Innovative research seeks to overcome these limitations, by leveraging novel techniques in temporal encoding and employing step-index few-mode fibers (FMFs), and enhance the computational capabilities of analog systems. In this work, we consider a 13 m FMF to implement up to 7-bit header classification at 28.5 Gb/s, when utilizing the information that is included in only one-bit duration. By activating a small number of spatial modes with proper dispersive optical characteristics, we are able to limit the complexity of the spatial transformation and solve such computing tasks efficiently. The selective activation of supported modes is performed by testing varying input displacement conditions between the optical beam encoding information and the FMF input tip. At the output of the fiber, we obtain multiple time series that correspond to different spatial sub-patterns of the output beam overall pattern. Finally, we train a logistic regression classifier that uses samples from these time series, and evaluate its performance in classification tasks. Notably, even in the presence of inter-header interference, our system achieves successful 6-bit header classification - a feat unattainable with conventional step-index MMFs.

In industry, for superior utilization of large capital equipment, a single large-current high-power converter is desired to convert electrical energy to a controlled high-intensity heat source to cater to the dynamic requirement of several welding methods (e.g., welding, hard-facing, gouging, weld overlay, etc.). The dynamic process behavior of multi-input minimally semiautomatic or single-input manual processes could follow either constant voltage (CV) or constant current (CC) arc type. Due to completely different type of arc or load characteristics, this article proposes a de-coupled control approach to guide the power source to independently take care of each arc type. It proposes feedforward control for CV arc type and robust second order sliding mode control (SOSMC) approach for dynamically uncertain CC arc type for manual welding. It further highlights that low-loss resonant or soft-switching topologies are less compatible for decoupled control of high-power loads with wide range impedance characteristics. For desired energy efficiency of a full-bridge DC-DC converter for arc welding, this article details that robustness features of SOSMC could be used for superior component engineering that incur low power loss. Finally, this article completely validates the proposed idea in a 1000A power converter for both CC and CV arc types.

Large-scale multi-building and multi-floor indoor localization has recently been the focus of intense research in indoor localization based on Wi-Fi fingerprinting. Although significant progress has been made in developing indoor localization algorithms, few studies are dedicated to the critical issues of using existing and constructing new Wi-Fi fingerprint databases, especially for large-scale multi-building and multi-floor indoor localization. In this paper, we first identify the challenges in using and constructing Wi-Fi fingerprint databases for largescale multi-building and multi-floor indoor localization and then provide our recommendations for those challenges based on a case study of the UJIIndoorLoc database, which is the most popular, publicly-available Wi-Fi fingerprint multi-building and multi-floor database. Through the case study, we investigate its statistical characteristics with a focus on the three aspects of (1) the properties of detected wireless access points, (2) the number, distribution, and quality of labels, and (3) the composition of the database records, and then identify potential issues and ways to address them in using the UJIIndoorLoc database. Based on the results from the case study, we not only provide valuable insights on the use of existing databases but also give important directions for the design and construction of new databases for large-scale multi-building and multi-floor indoor localization in the future.

This research addresses the pressing challenge of enhancing processing times and detection capabilities in Unmanned Aerial Vehicle (UAV)/drone imagery for global wildfire detection, despite limited datasets. Proposing a Segmented Neural Network (SegNet) selection approach, we focus on reducing feature maps to boost both time resolution and accuracy significantly advancing processing speeds and accuracy in real-time wildfire detection. This paper contributes to increased processing speeds enabling real-time detection capabilities for wildfire, increased detection accuracy of wildfire, and improved detection capabilities of early wildfire, through proposing a new direction for image classification of amorphous objects like fire, water, smoke, etc. Employing Convolutional Neural Networks (CNNs) for image classification, emphasizing on the reduction of irrelevant features vital for deep learning processes, especially in live feed data for fire detection. Amidst the complexity of live feed data in fire detection, our study emphasizes on image feed, highlighting the urgency to enhance real-time processing. Our proposed algorithm combats feature overload through segmentation, addressing challenges arising from diverse features like objects, colors, and textures. Notably, a delicate balance of feature map size and dataset adequacy is pivotal. Several research papers use smaller image sizes, compromising feature richness which necessitating a new approach. We illuminate the critical role of pixel density in retaining essential details, especially for early wildfire detection. By carefully selecting number of filters during training, we underscore the significance of higher pixel density for proper feature selection. The proposed SegNet approach is rigorously evaluated using real-world dataset obtained by a drone flight and compared to state-of-the-art literature.

In this paper, we investigate the performance of a Hybrid Quantum Neural Network (HQNN) and a comparable classical Convolution Neural Network (CNN) for detection and classification problem using a radar. Specifically, we take a fairly complex radar time-series model derived from electromagnetic theory, namely the Martin-Mulgrew model, that is used to simulate radar returns of objects with rotating blades, such as drones. We find that when that signal-to-noise ratio (SNR) is high, CNN outperforms the HQNN for detection and classification. However, in the low SNR regime (which is of greatest interest in practice) the performance of HQNN is found to be superior to that of the CNN of a similar architecture.

This paper describes an enhanced droop-based grid forming control method for inverter based distributed energy resources. The important advantage of inverter control over traditional synchronous machine behavior in terms of independent control of frequency and phase is highlighted. This degree of freedom is used to give the inverter control an inherent damped reaction to grid events, like voltage jumps, frequency or phase changes. The control is analyzed in the state space for three essential applications, i.e. grid parallel operation, two weak coupled power resources and two area system with two resources per area. It is shown, that the suitable damping coefficient mainly depends on the parameters of the power resource itself, like short circuit impedance and filter time constants, and is fairly independent on the residual system parameters. Simulation results confirm the analytical design of the damping coefficient. Experimental results gained in a complex power system laboratory validate the behavior for typical grid events.

The starting of decarbonization of electric power generation lead to an increase of renewable energy sources and a resulting reduction of conventional high-power plants. Thus, the inertia / stability of the grid is decreasing in many countries. In order to counteract this change grid forming control structures with virtual inertia for inverter based renewable energy sources have been developed in academia. In case of frequency and phase transients the frequency inertia of grid forming control will respond with grid supporting active power transients like synchronous generators would do. This is problematic for frequency and phase drops, if the regenerative power plant is already operating at its maximum active power available from the primary source. This paper describes effective measures to droop based grid forming control to avoid overloads at frequency or phase drops if there is no reserve power available. At the same time the presented measures do allow grid forming behavior to the extend active power reserve is available. Within this work an analytical solution for the power transient due to phase jumps is presented. Simulations and laboratory results with real world parameters confirm the effectiveness of the proposed measures and the coincidence with the theoretical results.

Polarimetric Synthetic Aperture Radar (PolSAR) images encompass valuable information that can facilitate extensive land cover interpretation and generate diverse output products. Extracting meaningful features from PolSAR data poses challenges distinct from those encountered in optical imagery. Deep Learning (DL) methods offer effective solutions for overcoming these challenges in PolSAR feature extraction. Convolutional Neural Networks (CNNs) play a crucial role in capturing PolSAR image characteristics by leveraging kernel capabilities to consider local information and the complexvalued nature of PolSAR data. In this study, a novel threebranch fusion of complex-valued CNN named Shallow to Deep Feature Fusion Network (SDF2Net) is proposed for PolSAR image classification. To validate the performance of the proposed method, classification results are compared against multiple stateof-the-art approaches using the Airborne Synthetic Aperture Radar (AIRSAR) datasets of Flevoland, San Francisco, and ESAR Oberpfaffenhofen dataset. The results indicate that the proposed approach demonstrates improvements in OA, with 1.3% and 0.8% enhancement for the AIRSAR datasets and 0.5% improvement for the ESAR dataset. The analyses conducted on the Flevoland data underscore the effectiveness of the SDF2Net model, revealing a promising OA of 96.01% even with only 1% sampling ratio. The source code is available at: https://github.com/mqalkhatib/SDF2Net

A document by Jinzhe Liu. Click on the document to view its contents.

This paper precisely evaluates the existing theoretical expressions for calculating guided power in symmetric dielectric slab waveguides for both transverse-electric (TE) and transverse-magnetic (TM) modes. In this work, we derive the relationships between electric and magnetic fields in both TE and TM modes by assessing potential values for propagation phase constant and subsequently assess the corresponding power expressions for determining total guided power flowing through the cross-section of symmetric dielectric slab waveguide. As a numerical example, we examine a typical dielectric slab waveguide consisting of a silicon core with a refractive index n1 = 3.5, surrounded by silicon dioxide cladding with refractive index n2 = 1.5, operating in optical frequency of 194 THz (i.e., λ ≈ 1.55 µm). Finally, we compare the calculated TE and TM mode field components and guided powers presented in figures and tables, respectively, validated by simulation results.

Quantitative assessment of spinal motion plays a pivotal role in diagnosing and understanding lower back pain. This paper utilises a Convolutional Neural Network for precise landmark localisation of bounding boxes encompassing the lumbar spine in sagittal plane lumbar fluoroscopy image sequences. The proposed methodology aims to automate spinal movement tracking and provide a benchmark for future research, thereby enhancing the efficiency and accuracy of low back pain diagnosis.

Astronomical object classification plays a crucial role in understanding the universe and its phenomena. In this study, the performance of pretrained convolutional neural network (CNN) architectures in classifying astronomical objects is evaluated using datasets from the Sloan Digital Sky Survey (SDSS). Five models—VGG19, ResNet101V2, ResNet152V2, DenseNet169, and DenseNet201—are analyzed for their accuracy in categorizing galaxies, stars, and quasars. The findings reveal ResNet152V2 as the top performer, achieving an accuracy of 92.09% and an AUC value of 0.989. These results underscore the significance of selecting appropriate pretrained models for accurate astronomical object classification.

In an era of significant economic volatility, time series forecasting is widely used to predict stock prices and guide investors in trading decisions. Nevertheless, existing data-driven techniques are unable to effectively handle the vast amount of financial data due to big data constraints such as nonlinearity, non-stationarity, heteroskedasticity, and unsynchronicity. A cohesive framework is also required for ensuring the smooth integration and synchronization of varied methodologies in timeseries financial prediction tasks. To address this problem, this paper introduces a novel framework that investigates three ensemble strategies: blending, stacking, and voting, and selects the best method to perform the stock trend prediction task. Specifically, we deploy four distinct machine learning algorithms as the base learning model, each of which is uncorrelated and proficient in a different way depending on the task. The outputs of the basis classifiers are then combined using the adaptive boosting algorithm, a meta classifier, to give the final prediction results. To augment predictive models's accuracy and generalization capabilities, we put forward strategies like feature engineering and Ridge regularization, which optimize the pertinence of data and curb overfitting. Our examination of five distinct case studies on Toronto Stock Exchange data reveals that the proposed multimodel ensemble method has superior performance compared to others.

This study introduces an innovative method for forecasting corn yield and price variations, critical for food security and strategic planning. Unlike traditional methods reliant on pre-harvest production data, which are often difficult to access, our approach utilizes satellite-derived gross primary production (GPP) data and dimension-reduction techniques to predict national corn yield and price changes. We conducted case studies in the US, Malawi, and South Africa to validate this approach, extracting predictors from annual GPP variations during peak growing seasons. We utilized various dimension reduction strategies, including spatial averaging, Empirical Orthogonal Functions (EOFs), and deep learning approaches like Autoencoders (AEs) and Variational Autoencoders (VAEs), to extract meaningful features from the GPP datasets. These extracted features serve as predictors in statistical models such as Generalized Linear Models and the Least Absolute Shrinkage and Selection Operator (LASSO), along with a neural network trained to predict variations directly from GPP-derived latent features. Model performances were evaluated using the Area Under Curve, Brier Skill Score, and Matthew Correlation Coefficient. Our results indicate that neural network models based on both AEs and VAEs exhibit superior predictive capabilities across all three countries, with the VAE excelling in the US and the AE leading in South Africa and Malawi. Similarly, the VAE outperforms other approaches for price prediction in the US and South Africa, while the AE achieves the best results in Malawi. Our study demonstrates the potential of combining satellite data and dimension reduction methods to significantly enhance large-scale corn yield and price predictions months before harvest.

This paper presents a broadband low-loss fanin wafer level ball grid array (WLB) vertical through-moldvia (TMV) interconnect that enables highly integrated systemin-package (SiP) applications beyond 220 GHz. The dedicate advantage of the proposed approach is to further minimize the separation between package components (e.g., antenna in package (AiP)) and on-chip receiver or transmitter chain. To demonstrate the robustness of the TMV interconnect design, the necessity for co-simulation of the integrated circuit (IC) package and the IC itself during the package-design is shown. Therefore, an in-depth analysis of the suppression of parasitic mode oscillation inside silicon is carried out. Furthermore, the parasitic radiation loss at the TMV interconnect discontinuity is given and a verification of the proposed TMV due to high agreement of simulation and measurement results is demonstrated. The interconnect has been verified with a measured 10 dB return loss bandwidth of 38 GHz and a de-embedded insertion loss of less than 2.8 dB at 275 GHz, which enables a compact low-loss broadband signal transition from active IC components in the backend-of-line (BEOL) to passive components in package. Compared to fan-out WLB, the developed fanin WLB solution enables higher integration density of active and passive components with the shortest interconnects for minimal insertion loss. Thus, the proposed fan-in TMV packaging provides a very compact, easy-to-assemble and cost-efficient alternative for SiP designs for future broadband communication and sensing solutions.