This study compares novel and established feature extraction techniques for classifying leaf diseases. Conventional methods like Transfer Learning and DCT are contrasted with novel techniques like Segmentation using a hetero- geneous data set of healthy and diseased leaf images. With little training data, transfer learning is beneficial, but with large data, more efficient al- gorithms may be more suitable, such as the custom segmentation model described in this work. The development of efficient plant disease detection systems using the information from these discoveries would improve agricul- tural practises and increase food security.

Precision agriculture brings together most innovative topics from research in the fields of vision, artificial intelligence and robotics aimed at developing sustainable and resilient solutions. In this context, crop identification is a crucial issue. This paper exposes a new algorithm to enhance real-time crop/weed instance segmentation, merging a learning-based method for instance segmentation and a feature model-based image processing strategy that relies on the vegetation characteristics of the crops. The proposed algorithm compensates for the shortcomings of each method performing independently. That is, the image processing strategy fulfils crop segmentation by generating finely refined masks; however, some weeds are segmented incorrectly. On the other hand, instance segmentation methods perform well on weeds and crops but, when inaccurate bounding boxes are present, imperfect masks are generated. The experiments are conducted in different evaluation campaigns within the ACRE international competition framework.

This paper introduces PIXIE (Penalised Pixel Intersection Error), a novel loss function aimed at improving model performance through the penalization of error pixels and alignment of the predicted mask with the ground truth mask. We explore the effectiveness of PIXIE alongside U-Net and DeepLabV3 networks. In addition to novel loss functions, this paper conducts model selection studies to understand overestimations and underestimations. We conduct a comprehensive evaluation using diverse datasets spanning medical imaging (breast tumor, COVID-19, Brain MRI) and remote sensing domains (forest fire, water bodies satellite images). Performance metrics, such as Intersection over Union (IoU), Dice Coefficient (DC), Area Error Ratio (AER), precision, and recall are quantitatively assessed. Moreover, we compare PIXIE with traditional loss functions like Jacard loss, Focal loss, and Binary crossentropy. PIXIE demonstrates comparable performance to the traditional loss functions in certain metrics and outperforms them in others, establishing itself as a leading approach in achieving exceptional results across all examined segmentation datasets. These findings represent a substantial contribution to the practical evolution of semantic segmentation in computer vision, offering essential insights into the optimization of loss functions for the development of accurate and robust models.

LCL filters with passive damping are widely used for grid-connected inverter (GCI) applications. This work proposes a novel non-iterative component selection procedure for the LCL filter with passive RC damping. While the primary goal of the filter is to effectively attenuate the switching frequency components of the inverter, there are several other desirable constraints that make the selection procedure challenging. In this work, eight such desirable constraints are identified from the existing literature and the proposed method meets all the identified constraints. Besides being non-iterative, the novelty of the work primarily lies in the selection of the damping resistor Rd and the ratio n in which the total shunt capacitor is split. The primary intent of Rd being to damp the resonance, a methodical approach that ensures a high impact of the resistance at the resonance frequency is proposed. The proposed approach also ascertains a minimum possible quality factor. The selection of the ratio n is leveraged as an additional degree of freedom. An analytical method to find the range for n that offers a critically flat resonance peak is proposed. A critically flat resonance peak is important for control stability. The final value of n is chosen from the available range to meet all the constraints.

Most existing methods for Magnetic Resonance Imaging (MRI) reconstruction with deep learning use fully supervised training, which assumes that a high signal-to-noise ratio (SNR), fully sampled dataset is available for training. In many circumstances, however, such a dataset is highly impractical or even technically infeasible to acquire. Recently, a number of self-supervised methods for MR reconstruction have been proposed, which use sub-sampled data only. However, the majority of such methods, such as Self-Supervised Learning via Data Undersampling (SSDU), are susceptible to reconstruction errors arising from noise in the measured data. In response, we propose Robust SSDU, which provably recovers clean images from noisy, sub-sampled training data. Robust SSDU trains the reconstruction network to map from a further noisy and sub-sampled  version of the data to the original, singly noisy and sub-sampled data, and applies a Noisier2Noise correction term at inference. We also present a related method, Noiser2Full, that recovers clean images when noisy, fully sampled data is available for training. Both proposed methods are applicable to any network architecture, straight-forward to implement and have similar computational cost to standard training.  We evaluate our methods on the multi-coil fastMRI brain dataset with a novel denoising-specific architecture and find that it performs competitively with a benchmark trained on clean, fully sampled data.

The paper advocates for a dual approach to software documentation that amalgamates mathematical notation and flowcharts. This approach aims to enhance both accessibility and quality control. Mathematical notations allow for precise understanding, while flowcharts provide a comprehensive view of complex systems. These components collectively broaden the research's appeal to a multidisciplinary audience and impose a level of scrutiny that augments reliability. The framework further bolsters code maintainability and extendability, contending that rudimentary inline comments are inadequate for thorough documentation. An illustrative example using deconvolution is included.

This paper was published in IEEE   Access, vol. 12, pp. 125942-125954, 2024, doi: 10.1109/ACCESS.2024.3409738.TWe propose a visual programming framework that helps a designer easily convert an existing analog layout into the layout generator. Using a graphical user interface (GUI), designers can easily load an existing layout, convert it into the layout generator, and visually verify the generated layout result. A GUI-supported visual programming method enables intuitive and straightforward programming to significantly reduce the required programming skills and coding workload. Through program blocks, designers can easily describe and compile a layout generator. Layout-code synchronization updates the program blocks automatically when layout elements are created, edited, or deleted via GUI. Expression assistance semi-automatically completes parameterized expressions of layout geometry through simple GUI manipulation. These methods greatly reduce the time and workload for development of the analog layout generator. Using the framework, a complex 20 Gb/s high-speed receiver was converted to the generator in 20 hours, which is just 45% of the time required for manual coding of the layout generator. Other high-speed analog blocks including DCDL, and DCC circuits were also converted to generators. The generators created hundreds of different layout designs, all of them passed design rule verification. The generated layouts achieved almost the same performance in post-layout simulation as the reference layouts.

Use of fingerprints for identity in digital world arose very quickly but the smartphones' sensor don't provide such services for cloud based authentications due to TEE policy of manufacturers. In this paper we developed a novel technique to use smartphones' sensor for fingerprints data storage and authentication purpose.

In the context of energy transmission, Line Rating addresses the capacity of conductors to safely transmit energy under different conditions. Two distinct rating methods are known: Static Line Rating and Dynamic Line Rating, cor?responding to fixed worst-case weather conditions and real?time variable conditions, respectively. The assumption underlying Static Line Rating, that conductor temperature and line sag are deterministically correlated, is challenged by real-world findings, revealing stochastic relationships influenced by uncontrollable weather variability. The potential for safely operating trans?mission lines beyond their static-rated capacity is discussed, emphasizing the use of Dynamic Line Rating with advanced technologies. A novel architecture involving temperature sensors, laser diodes, and optical fibers is introduced for achieving Dynamic Line Rating. The proposed system enables real-time temperature monitoring and transmission to a control station for optimized energy transport. Various implementation options, from integrated solutions to external attachments, are considered. The abstract concludes with an overview of the learning process for modeling temperature behaviors and line current allocation using an Echo State Network, showcasing its suitability for capturing dynamic complexities in the context of transmission line rating.

We use the following anomaly detection benchmark data sets: KDDCUP is an old cyber-intrusion detection benchmark data set and conveys some well-known issues such as duplicated data points. It contains simulated military-like data for normal traffics and several types of attacks. NSL-KDD is a revisited version of the KDDCUP data set provided by the Canadian Institute of Cybersecurity. CIC-CSE-IDS2018 is a recent simulated data set containing normal traffics and several types of attacks from a complex network. It is a collaboration-work production between the Canadian Institute of Cybersecurity and the Communication Security Establishment [57]. \tightlist

Fog computing reduces network latency by moving computational resources close to where the data is generated. Vehicular fog computing (VFC) is an emerging computing paradigm where fog nodes deployed on moving vehicles (i.e., vehicular fog nodes (VFNs)) complement stationary fog nodes (e.g., the ones co-located with cellular base stations) to satisfy the spatiotemporally varying demand for computing resources in a cost-efficient manner. On-demand VFC (ODVFC) supports dynamic routing of VFNs, with the aim of fulfilling the spatiotemporally varying demand for computational resources in a cost-efficient manner. Different from previous works on capacity planning and vehicle routing that utilize compute-intensive optimization methods such as integer linear programming (ILP), this paper explores the feasibility of applying reinforcement learning to dynamic capacity planning in a time-efficient manner. Specifically, we propose to apply multi-agent reinforcement learning (MARL) with actor-critic methods to train the VFN routing policies. This approach allows distributed VFNs to cooperatively maximize the techno-economic performance of ODVFC. For evaluation, we built an open-source VFC simulation platform that integrates vehicular traffic simulation with 5G NR V2X and MARL environment. Compared with decentralized learning (i.e., each VFN independently learns its routing policy), centralized learning (i.e., using a global agent for VFN routing), and ILP methods, our proposal proves to achieve 8.3% higher revenue and 13.2% higher number of served tasks than decentralized training; and it has 40.6% and 83% lower execution time than centralized learning and ILP, respectively, with only 14% lower revenue than both. It is also scalable to real-life scenarios with a great number of users and VFNs.

Abstract: Gait is an indicator of a person’s health status and abnormal gait patterns are associated with a higher risk of falls, dementia, and mental health disorders. Wearable sensors facilitate long-term assessment of walking in the user’s home environment. Earables, wearable sensors that are worn at the ear, are gaining popularity for digital health assessments because they are unobtrusive and can easily be integrated into the user’s daily routine, for example, in hearing aids. A comprehensive gait analysis pipeline for an ear-worn accelerometer that includes spatial-temporal parameters is currently not existing. Therefore, we propose and compare three algorithmic approaches to estimate step length and gait speed based on ear-worn accelerometer data: a biomechanical model, feature-based machine learn- ing (ML) models, and a convolutional neural network. We evaluated their performance on a step and walking bout level and compared it with an optical motion capture system. The feature-based ML model achieved the best performance with a precision of 4.8 cm on a walking bout level. For gait speed, the machine learning approach achieved an absolute percentage error of 5.3% (± 3.9%). We find that the ML model is able to estimate step length and gait speed with clinically relevant precision. Furthermore, the model is insensitive to different age groups and sampling rates but sensitive to walking speed. To our knowledge, this work is the first contribution to estimating step length and gait speed using ear-worn accelerometers. Moreover, it lays the foundation for a comprehensive gait analysis framework for ear-worn sensors enabling continuous and long-term monitoring at home.

We consider one essential performance metric of caching schemes based on two different coding techniques. A heterogeneous cache network is assumed, in which a server is connected through a shared link to two relays while there are not connection between the server and users. The number of users connected to each relay has a Poisson distribution. The memories in the first scheme are filled with encoded data while they are filled with uncoded data in the second scheme, and encoding of the content is utilized during the delivery phase. Knowing the popularity distribution of the files and the number of users, the goal is to calculate the average transmission rate of the shared link. However, we show here that this metric can be calculated in the general case of having a random number of users. In particular, a fundamentally different approach is needed, in which we employ the results of balls and bins (BAB) as a tool to solve these problems. We present firstly that this mapping has the advantage of an easy computation for an uncoded caching. We compare the schemes by characterizing the average transmission rate when the demands of users have the uniform and nonuniform distribution.

The COVID-19 pandemic is an unparalleled global crisis that has led to a surge in innovative responses. Central to this innovation landscape was the emergence of open-source projects that offered rapid, collaborative solutions, complementing scientifically led endeavors. This paper undertakes a comprehensive exploration of open-source innovations during the pandemic, chronicling their instigation, impact, and broader implications for research and development (R&D) paradigms. Utilizing a meticulously curated database of open-source COVID-19 innovations, we compared these projects with scientifically led initiatives and derived insights into their respective strengths, challenges, and contributions. Our findings underscore the transformative potential of collaborative research models, emphasizing the synergies between open source and scientific approaches. Furthermore, we delineate the broader implications of these findings for future R&D landscapes, highlighting the pivotal roles of community engagement, adaptability, and interdisciplinary collaboration. This study offers a holistic understanding of innovation dynamics during global crises, serving as a guide for future research endeavors and policy formulations.

The utilization and engineering of thermo-optic effects have found broad applications in integrated photonic devices, facilitating efficient light manipulation to achieve various functionalities. Here, we perform both an experimental characterization and theoretical analysis of these effects in integrated micro-ring resonators made from high index doped silica (HIDS), which have had many applications in integrated photonics and nonlinear optics. By fitting the experimental results with theory, we obtain fundamental parameters that characterize their thermo-optic performance, including the thermo-optic coefficient, the efficiency for the optically induced thermo-optic process, and the thermal conductivity. The characteristics of these parameters are compared to those of other materials commonly used for integrated photonic platforms, such as silicon, silicon nitride, and silica. These results offer a comprehensive insight into the thermo-optic properties of HIDS based devices. Understanding these properties is essential for efficiently controlling and engineering them in many practical applications.

Demand for autonomous protection in computing devices can not go unnoticed with a cataclysmic rise in cyber-attacks. Consequently, cybersecurity measures with an improved generalization that can proactively determine the indicators of compromises to predict zero-day threats or previously unseen malware together with known malware are highly desirable. In this article, we present a novel concept of autonomous device protection based on behavioral profiling by continuously monitoring internal resource usage and exploiting a large language model to distinguish between benign and malicious behavior. We design and develop a proof-of-concept for Windows-based computing devices relying on a built-in event tracing mechanism for log collection that is converted into structured data using a graph data structure. We extract graph-level features, \textit{i.e., graph depth, nodes count, number of leaf nodes, node degree statistics, and events count}, and node-level features, i.e., process start, file create and registry events details for each graph. Further, we exploit a pre-trained large language network - a simple contrastive sentence embedding framework to extract strong features, i.e., dense vectors, from event graphs. Finally, we train a random forest classifier using both the graph- and node-level features to obtain classification models that are evaluated on a collected dataset containing one thousand benign and malicious samples achieving accuracy up to 99.25%.

In industrial settings, robots are typically employed to accurately track a reference force to exert on the surrounding environment to complete interaction tasks. Interaction controllers are typically used to achieve this goal. Still, they either require manual tuning, which demands a significant amount of time, or exact modeling of the environment the robot will interact with, thus possibly failing during the actual application. A significant advancement in this area would be a high-performance force controller that does not need operator calibration and is quick to be deployed in any scenario. With this aim, this paper proposes an Actor-Critic Model Predictive Force Controller (ACMPFC), which outputs the optimal setpoint to follow in order to guarantee force tracking, computed by continuously trained neural networks. This strategy is an extension of a reinforcement learning-based one, born in the context of human-robot collaboration, suitably adapted to robot-environment interaction. We validate the ACMPFC in a real-case scenario featuring a Franka Emika Panda robot.Compared with a base force controller and a learning-based approach, the proposed controller yields a reduction of the force tracking MSE, attaining fast convergence: with respect to the base force controller, ACMPFC reduces the MSE by a factor of 4.35.

In the context of mobile and Internet of Things (IoT) networks, data naturally originates at the edge, making crowdsourcing a convenient and inherent approach to data collection. However, crowdsourcing presents challenges related to privacy, sampling bias, statistical sufficiency, and the need for time-consuming post-processing. To this end, generating synthetic data using Deep Learning techniques emerges as a promising solution to overcome such limitations. In this study, we propose an innovative framework that transcends applications and data types, enabling the conditional generation of crowdsourced datasets with location information in mobile and IoT networks. A crucial aspect of our methodology is its ability to assess uncertainty in newly generated samples and produce calibrated predictions through approximate Bayesian methods. Without loss of generality, we ascertain the validity of our method on the task of Minimization of Drive Test (MDT) data generation, presenting for the first time a comparison of synthetically generated data with an original large-scale MDT set collected from a Mobile Network Operator’s network infrastructure. By offering a versatile solution to data generation, our framework contributes to overcoming challenges associated with crowdsourced data, opening up possibilities for advanced analytics and experimentation in mobile and IoT networks.