GSP-enhanced embedded hardware framework for real-time multi-class road target detection in ADAS using mmWave radar

Facial expression is significant for face-to-face communication since it is one of our body language that increases data information during the communication. In recent surveys, it found that feature extraction methods have influenced on facial recognition directly and they can outperform if some irrelevant features are eliminated. Consequently, a procedure of facial expression recognition using graph-based features and artificial neural network is proposed in this thesis, and the procedure can be divided into 2 phases. For the first phase, fourteen points are manually located to create graph with edges connecting among such points, followed by computation of the Euclidean distances from those edges to define them as features for training in the next phase. For the next phase, Multilayer Perceptrons with back-propagation learning algorithm is implemented to recognize six basic emotions from the corresponding feature vectors. To evaluate the performance, Cohn-Kanade AU-Coded facial expression database is applied to the recognition system under various kinds of feature and training technique, and the experimental results have shown that MLP (Graph-based features) without validation and MLP (Graph-based feature) with cross validation can achieve the highest recognition rate (95.24%). Therefore, it has illustrated that the combination of graph-based features and artificial neural network is the efficient way for the facial expression recognition.

Peer-to-Peer (P2P) lending is an online platform to facilitate borrowing and investment transactions. A central problem for these P2P platforms is how to identify the most influential factors that are closely related to the credit risks. This problem is inherently complex due to the various forms of risks and the numerous influencing factors involved. Moreover, raw data of P2P lending are often high-dimension, highly correlated and unstable, making the problem more untractable by traditional statistical and machine learning approaches. To address these problems, we develop a novel filter-based feature selection method for P2P lending analysis. Unlike most traditional feature selection methods that use vectorial features, the proposed method is based on graph-based features and thus incorporates the relationships between pairwise feature samples into the feature selection process. Since the graph-based features are by nature completed weighted graphs, we use the steady state random walk to encapsulate the main characteristics of the graph-based features. Specifically, we compute a probability distribution of the walk visiting the vertices. Furthermore, we measure the discriminant power of each graph-based feature with respect to the target feature, through the Jensen-Shannon divergence measure between the probability distributions from the random walks. We select an optimal subset of features based on the most relevant graph-based features, through the Jensen-Shannon divergence measure. Unlike most existing state-of-the-art feature selection methods, the proposed method can accommodate both continuous and discrete target features. Experiments demonstrate the effectiveness and usefulness of the proposed feature selection algorithm on the problem of P2P lending platforms in China.

Liver disease diagnosis remains challenging due to asymptomatic early stages and limitations of traditional diagnostic methods. This paper presents a novel graph-theoretic approach combined with advanced data balancing for improved liver disease classification. We develop a patient similarity graph using cosine similarity of biochemical features, extracting five centrality measures (degree, clustering coefficient, betweenness, closeness, and eigenvector centrality) to capture relational patterns overlooked by conventional methods. The Indian Liver Patient Dataset is enhanced from 10 to 15 features through graph-based feature engineering. Class imbalance is addressed using SMOTEENN technique. An ensemble voting classifier incorporating multiple algorithms (XGBoost, LightGBM, Random Forest, etc.) is evaluated via 10-fold cross-validation. Our approach achieves superior performance with 95.74% accuracy, 94.25% precision, 99.31% recall, 96.67% F1-score, and 98.53% ROC AUC, significantly outperforming existing methods. Results demonstrate that leveraging patient relationships through graph-based features substantially enhances diagnosis accuracy.

Joint analysis of crash injury severities for autonomous and conventional vehicles in mixed traffic environments: Application of random parameter bivariate probit model.

Automated monitoring of Personal Protective Equipment (PPE) is crucial for enhancing safety in high-risk environments like construction sites, yet selecting the optimal detection model requires careful evaluation of accuracy versus efficiency trade-offs. This study presents a comparative performance analysis across distinct object detection paradigms represented by YOLO (YOLOv8, YOLOv11n), Faster R-CNN, and DETR to benchmark their suitability for real-time PPE detection. However, this study moves beyond a simple technical benchmark by also proposing a logical process to transform raw model detections (e.g., 'person', 'hardhat') into actionable compliance verification information (e.g., 'Compliant'/'Non-Compliant'). Using a curated construction site safety dataset, models were evaluated based on standard accuracy metrics (including mAP@.5:.95) and efficiency measures (inference latency). Results indicate that DETR and YOLOv11n achieved the highest overall accuracy with an identical mAP@.5:.95 of 0.770, closely followed by YOLOv8 (0.763), while the YOLO family demonstrated significantly superior real-time efficiency (6-7 ms latency). Faster R-CNN recorded a lower mAP (0.703) and the highest latency. Conclusively, YOLOv11n offers the most compelling balance for the detection phase, and the proposed logical process provides a practical method for integrating this technical output into automated safety monitoring systems.

Accurate prediction of molecular toxicity is vital for drug development. Most mainstream methods rely on fingerprints or graph-based feature extraction, the emergence of large language models (LLMs) offers new prospects for molecular representation learning in toxicity prediction. Although several studies attempt to leverage LLMs to integrate molecular sequence data for pretraining molecular representations, certain limitations remain. Current LLM-based approaches usually utilize solely on class embedding features, overlooking the rich information in sequence embedding. Moreover, integrating pre-trained molecular representations with multi-modal molecular data may further enhance performance in toxicity prediction. To address these challenges, we propose BFGTP, a BERT-guided two-stage molecular representation learning framework for toxicity prediction. Firstly, we design independent encoders for molecular descriptions of three modalities, where the fingerprint encoder with dual level attention mechanisms effectively integrates multi-category fingerprints. Then, the two-stage guide strategy is introduced to fully utilize the prior knowledge of LLMs, employing contrastive learning to align and fuse the tri-modal representations and knowledge distillation to align predicted value distributions. BFGTP ultimately combines fingerprint and graph representations to predict molecular toxicity. Experiments on seven toxicity datasets show that BFGTP outperforms baselines, achieving the highest AUC on five datasets and the best average performance across five evaluation metrics. Ablation studies, t-SNE visualization and case study confirm the effectiveness of BFGTP's components and its ability to capture meaningful molecular representations.

Reliable road object detection is crucial for self-driving cars and intelligent transportation systems. However, vision-only techniques fail in low-visibility conditions such as fog, rain, or nightfall, whereas deep radar-based models frequently suffer from high latency and energy consumption. To overcome these restrictions, we propose a millimeter-wave (mmWave) radar-based object identification system based on a PYNQ-ZU (Field-Programmable Gate Array) FPGA that achieves a compromise between accuracy, efficiency, and real-time performance. Our proposed pipeline transforms 3D radar point clouds to 2D Top-View (TV) representations and uses the Spectral Graph Wavelet Transform (SGWT) to extract discriminative spatialtemporal features with little computational cost. A Random Forest (RF) classifier achieves 97% accuracy across seven object classes with an end-to-end latency of 61 ms. When it comes to power performance, the FPGA implementation uses 119.32 mW at idle, 268.48 mW on average, and 357.97 mW at its peak, demonstrating its energy efficiency under dynamic workloads. Compared with deep learning alternatives like Tiny Convolutional Neural Network (CNN) (73.60 MB, 4493 mW, 449 ms), the model size is 6.47 MB, which is substantially smaller and more power-efficient. By overcoming the limitations of vision-only systems (poor visibility) and deep radar models (high latency, energy, and memory), the proposed SGWT + RF framework enables real-time, multi-class detection in resource-constrained environments—offering a practical and robust solution for autonomous navigation and traffic monitoring applications.

Nonlinear fibre propagation and transceiver impairments limit the performance of coherent optical communication systems. Established digital signal processing techniques such as digital back-propagation and Volterra equalisation can mitigate Kerr-induced nonlinear distortions, but their computational complexity and energy consumption hinder real-time deployment. This thesis develops machine-learning and neuromorphic approaches to optical channel equalisation, focusingon performance–complexity trade-offs and practical hardware feasibility. A unified training and benchmarking framework is introduced for coherent transmission scenarios, employing Q-factor and bit-error-rate alongside implementation measures including operation counts, memory footprint, and inference latency. Within this framework, a broad class of equalisers is analysed, including multi-layer perceptrons, convolutional and recurrent neural networks, and complex-valued neural networks that process in-phase and quadrature components. Model compression is studied: pruning, quantisation, and weight clustering are jointly optimised using Bayesian optimisation to identify Pareto-efficient configurations that preserve equalisation performance while reducing computational load, memory usage, and latency. Experimental evaluations on edge-device platforms demonstrate feasibility under realistic receiver constraints. The thesis also explores hardware–software co-design. Optical phase conjugation is integrated with neural equalisers to offload part of the nonlinearity compensation to the optical domain, enabling smaller models and lower digital complexity. To address analogue noise in photonic neuromorphic computing, robustness is assessed under additive and signal-dependent noise; noise-aware training, stochastic-resonance neurons, and ensemble-based “crowd equalisation” are proposed to enhance resilience. Finally, a neuromorphic equaliser combining spiking neural networks with a streaming RWKV time-mixing module is introduced. By exploiting event-driven sparsity and constant-memory sequential processing, this architecture achieves competitive equalisation performance with reduced computational and energy requirements compared to conventional deep learning models.

Bridging the communication gap for Deaf communities in Indonesia requires recognizing everyday social-command gestures beyond alphabet-focused signs. We present an end-to-end deep-learning framework for real-time recognition of 19 common Indonesian "silahkan" gestures (e.g., "silahkan makan" [please eat], "silahkan belajar" [please study], "silahkan berdoa" [please pray]) from live webcam video. Each 16-frame clip is cropped to the hand region using MediaPipe Hands and processed by a shared MobileNetV2 backbone. We compare two lightweight classification heads: a CNN-only model that achieves 93.21% accuracy (macro-F = 0.9289) with 3.8 ms inference latency per sample, and a CNN+LSTM model yielding 93.83% accuracy (macro-F = 0.9386) with 9.4 ms latency. Paired t-tests confirm the significance of these differences (p = 0.0003). Blind trials on 56 novel videos further demonstrate robustness—CNN-only achieves 98.21% accuracy while CNN+LSTM reaches 89.29%. Our entirely CPU-based approach enables efficient, accessible gesture recognition for Indonesian Deaf communities.

In this paper, we propose a deep learning pipeline for real-time crop disease classification on mobile devices. Our system employs a custom Convolutional Neural Network (CNN) trained on publicly available crop disease datasets (Maize, Tomato, Potato, Rice). In addition, two transfer-learning models; ResNet-50 and MobileNet are used as fixed feature extractors, with their output features classified by a multi-class Support Vector Machine (SVM) with Radial Basis Function (RBF) kernel. We compare the models’ performance across all crop datasets and evaluate inference latency and model size. Experimental results show that the ResNet50-SVM hybrid attains near-perfect accuracy (≈100% for Maize, Tomato, Potato; 99.96% for Rice) on plant disease classification, far exceeding both the custom CNN and MobileNet-SVM approaches. The MobileNet-SVM pipeline is notably faster (≈23–66 ms per image) and compact (~8.7 MB) than ResNet50+SVM (≈108–192 ms, ~90 MB), making it well-suited for on-device deployment. The final model is converted to TensorFlow Lite for mobile inference; on a typical smartphone CPU it processes an input image in ~0.15–0.19s on average, enabling practical field use. These results demonstrate an efficient mobile AI solution for crop disease detection that balances accuracy with resource constraints. The proposed system can empower farmers with timely, in-field disease diagnosis, helping to mitigate yield losses and improve crop management through accessible AI-driven tools.

A hybrid wired/wireless SDN switch enables unified control across mixed-access networks but faces challenges in real-time traffic classification and flow management due to the lack of intelligence. This paper presents an intelligent hybrid SDN switch augmented with embedded machine learning (ML) models for accurate traffic classification and low-latency prediction. We evaluated six supervised ML algorithms, Logistic Regression, Random Forest, Support Vector Machine, Naive Bayes, k-Nearest Neighbors, and Neural Network, under both single-flow and concurrent multiflow scenarios using real traffic traces, including DNS, Ping, Telnet, and interactive gaming protocols. The evaluation results show that the Random Forest model consistently outperforms others, achieving classification accuracies of 99.97% in single-flow and 99.86% in multi-flow scenarios, while accurately predicting the traffic in millisecond-level inference latency. These findings demonstrate the feasibility of embedding ML-driven intelligence into hybrid SDN switches to support real-time, application-aware traffic control in emerging network infrastructures.

The installation of surveillance systems in public or private settings is increasingly integrated with intelligent and real-time video anomaly detection to proactive threat consideration. A framework is proposed for real-time abnormal behavior detection for streaming video data based on Dynamic Graph Neural Networks (DGNNs). Unlike traditional frame-based models, the DGNN is not restricted to a static set of nodes; instead, it treats each frame as a graph, where nodes denote detected physical objects such as people and vehicles. DGNN introduces real-time frame-wise graph updates unlike prior methods with static or slow graph refresh using motion, pose, and interaction cues, supported by a lightweight spatio-temporal GCN optimized for live anomaly detection. The results convey high performance for the model, which recorded an accuracy of 99.65%, along with the corresponding precision, recall, and F1-score. An FNR and FPR of only 0.35% and 0.03%, respectively, highlighted its strong performance in discriminating normal from anomalous behaviors. The ROC and precision-recall AUC scores zoomed close to 1.0000 in almost all categories, thereby attesting to the system’s robustness where even slight stakes are involved. Following this was the fusion of TensorRT optimization in ensuring the real-time processing capability of the model, with an inference latency of fewer than 200 milliseconds per segment. The DGNN system, therefore, represents a scalable, low-latency option that is capable of real-time detection and classification of diverse and complex real-world abnormal behaviors, making it a novel approach toward smart public safety systems with very little dependency on huge annotated datasets.

A Supervised Contrastive Learning Framework with Graph-Based Feature Extraction for Small-Sample Automatic Modulation Recognition

The growing reliance on natural gas in urban China has heightened the urgency of maintaining pipeline integrity, particularly in environments prone to disruption by nearby construction activities. In this study, we present a practical approach for the real-time classification of distributed fiber optic monitoring signals, leveraging a hybrid framework that combines the feature learning capacity of a one-dimensional convolutional neural network (1D-CNN) with the classification robustness of a support vector machine (SVM). The proposed method effectively distinguishes various pipeline-related events—such as minor leakage, theft attempts, and human movement—by automatically extracting their vibration patterns. Notably, it addresses the common shortcomings of softmax-based classifiers in small-sample scenarios. When tested on a real-world dataset collected via the DAS3000 system from Hangzhou Optosensing Co., Ltd., the model achieved a high classification accuracy of 99.92% across six event types, with an average inference latency of just 0.819 milliseconds per signal. These results demonstrate its strong potential for rapid anomaly detection in pipeline systems. Beyond technical performance, the method offers three practical benefits: it integrates well with current monitoring infrastructures, significantly reduces manual inspection workloads, and provides early warnings for potential pipeline threats. Overall, this work lays the groundwork for a scalable, machine learning-enhanced solution aimed at ensuring the operational safety of critical energy assets.