Medication errors remain a critical challenge in healthcare, particularly for elderly patients and visually impaired individuals. Although existing pill image recognition systems show promise, most operate under controlled laboratory conditions, limiting their real-world applicability. This paper presents an efficient framework for real-time pill image recognition on edge devices using Adaptive Lightweight Attention (ALA). Our approach combines multimodal feature extraction from RGB, contour, texture, and text modalities with an innovative ALA fusion mechanism to achieve high accuracy while maintaining computational efficiency. We conduct dual-dataset evaluation to demonstrate framework adaptability: on the challenging CURE dataset (196 classes, 20–62 samples/class) with cluttered backgrounds and uncontrolled illumination, ALA achieves 96.23

This paper proposes AFB-Net, an anomaly-based timeframe for screening for autism spectrum disorder (ASD). The model learns typical behavioral dynamics from typically developing (TD) subjects and identifies deviations as potential indicators of ASD. To enhance temporal representation, the Top-K anomaly segment selection strategy is used to focus on the most informative segments. Experimental results showed AUCs ranging from 0.63 to 0.66 across the feature modes. Head posture yielded the best results, achieving an accuracy of 0.67 and sensitivity of up to 0.90. Facial and eye features provided additional but weaker contributions. The model consistently exhibited high sensitivity but relatively low specificity, suggesting a tendency to detect ASD cases well. Time-based analysis revealed that ASD-related anomalies were sparse and concentrated in high-anoma segments, effectively supporting the Top-K strategy. Noise pulses in TD samples diminished the model's robustness. This finding highlights the potential of anomaly-based time-based modeling.

Smartphone-based colorimetric sensing offers a low-cost route for chemical quantification, but its reliability is challenged by device-dependent color responses and calibration requirements. This paper revisits whether explicit calibration remains useful when ammonium $\left( {{\mathbf{NH}}_4^ + } \right)$ concentration is estimated by image-native deep regression. Using a laboratory 10K benchmark containing 9,922 clean smartphone strip images, we evaluate a regression pipeline with group-aware partitioning, YOLO-based strip cropping, optional green-border mean normalization, and a deep regressor trained on cropped test-strip images. Five representative backbones, namely MobileNetV3-Large, EfficientNet-B0, TF-EfficientNet-B3, ConvNeXt-Tiny, and Swin-Tiny, are compared under two input settings: no explicit calibration and green-border mean normalization. Across all evaluated backbones, the uncalibrated setting outperforms green-border normalization for ${\text{NH}}_4^ +$ regression, with the best overall results obtained by TF-EfficientNet-B3 without explicit calibration. These findings indicate that, under the present deep-regression setting, the tested green-border normalization is unnecessary and may degrade predictive performance.

Spectral clustering, one of the most popular modern clustering algorithms, explores the connectivity between data points and optimizes the minimum graph cut on the graph, resulting in clusters in the dataset. Despite the ability to capture the geometric properties of a graph for clustering non-convex data, this algorithm relies on building an accurate initial affinity graph. Consequently, it suffers from a dependency on pre-defined parameters to correctly represent connections among data. Self-adaptive graph-based approaches have emerged to address this limitation, but they require capturing the local and global cluster information to achieve the optimal graph structure. To mitigate these problems, we introduce a self-adaptive graph clustering framework that dynamically adjusts the affinity graph structure and imposes several constraints, such as sparsity, matrix rank and node invariance. Moreover, to improve the quality of cluster results, our framework replaces traditional k-means clustering with a soft-label assignment module and optimizes it in a self-supervised manner. The experimental results demonstrate that our modifications can construct optimal graph structures and deliver remarkable improvements over the baseline model and the existing clustering methods.

Vietnamese visual speech recognition remains under-resourced, especially in compact deployment-oriented settings. Rather than attempt a full Vietnamese VSR benchmark, this paper asks whether a large timestamped Vietnamese lip corpus can be distilled into a smaller subset better aligned with Jetson-class local inference. Starting from the parent collection, which provides word-level timestamps and derived mouth crops, we mine unigram and bigram candidates, filter the vocabulary with frequency, diversity, duration, and visual-salience criteria, and enforce grouped, balanced splits. A PyTorch LipNet-style recognizer with a 3D convolutional front end, bidirectional recurrent modeling, and character-level CTC decoding is then trained under two label-space designs: mixed unigram+bigram and unigram-only. On the balanced test subsets, unigram-only improves exact match from 27.09% to 32.35% and reduces WER from 0.721 to 0.676 while keeping CER nearly unchanged (0.540 vs. 0.541). The findings indicate that, for compact Vietnamese lip reading, vocabulary design and leakage-aware subset construction are as important as model choice, and that a unigram-only label space is better suited to fixed-budget edge inference.

Visual speech recognition is a central problem in computer vision, encompassing both lip reading (visual speech recognition) and sign language recognition. Although substantial progress has been achieved independently on each task, their complementary characteristics have rarely been explored jointly. In this work we propose UniModal-LSR (Unified Multimodal Lip and Sign Recognition), a novel deep learning framework that jointly addresses lip reading and sign language recognition within a single multimodal architecture. By exploiting shared properties of visual communication channels, namely temporal dynamics, spatial articulation structure, and contextual dependencies, the proposed model enables bidirectional transfer of knowledge between modalities. The framework incorporates a Hierarchical Temporal-Spatial Encoder that captures multi-scale temporal patterns through the combination of local convolutions and global self-attention. It also includes a Cross-Modal Attention Fusion module that performs dynamic, context-aware information exchange via bidirectional cross-attention and adaptive gating. Additionally, a Contrastive Semantic Alignment loss enforces semantic consistency across modality-specific representations. Overall, the architecture integrates three-dimensional convolutional neural networks for spatiotemporal feature extraction with graph neural networks for explicit hand-pose modeling. Extensive experiments on several public benchmarks show that UniModal-LSR improves performance compared with recent methods. The model attains a Word Error Rate (WER) of 33.2% on LRS2-BBC, representing a 12.4% relative gain. On PHOENIX-2014, it achieves 18.3% WER, a 13.7% relative gain. Moreover, the unified model reduces parameter count by 25.9% relative to two separate task-specific systems. These results indicate that unified multimodal modeling can improve visual speech recognition performance and may support future communication technologies.

BACKGROUND:Rapid, low-cost detection of nutrient pollutants such as nitrite (NO2-) and ammonium (NH4+) is crucial for environmental monitoring. Conventional colorimetric instruments offer high accuracy but are expensive, bulky, and unsuitable for on-site analysis. Smartphone-based colorimetry provides a portable and affordable alternative, yet its accuracy is often compromised by variations in illumination and camera hardware. Although color reference standards and linear calibration improve consistency, achieving device-independent quantification remains challenging. This study introduces an integrated smartphone colorimetric platform combining a controlled lightbox, embedded color reference, and ensemble machine learning to enhance analytical robustness and reproducibility. RESULTS:A total of 2,700 Rhodamine B images (five smartphones, 30 concentrations each) and 5,400 analyte images (six smartphones, 30 concentrations each for NO2- and NH4+) were analyzed using eight machine learning algorithms under two pipelines: ROI-only and reference "Square" calibration. Ensemble models (Random Forest, XGBoost) consistently outperformed other approaches, achieving >95 % accuracy in concentration classification and superior regression performance. Reference calibration markedly enhanced cross-device reproducibility and precision, increasing R2 from 0.76-0.89 (ROI) to 0.89-0.95 and reducing MSE by 25-75 %, depending on model and device (e.g., Mi8 Lite-XGBoost: 2.16 → 0.57 ppm2; Nokia-Random Forest: 1.49 → 0.85 ppm2). Even budget smartphones (e.g., Oppo A83, Redmi A1) reached R2 ≥ 0.91, comparable to high-end models. Variance across replicate trials decreased, confirming improved stability. Combined RGB/HSV/CIELAB feature sets with reference-normalized deltas and ratios were most predictive, yielding a final R2 of 0.95 with minimal error and high consistency. SIGNIFICANCE:By integrating standardized imaging, color reference calibration, and ensemble machine learning, this study establishes a robust and device-independent framework for quantitative smartphone colorimetry. The significant improvements in predictive accuracy (higher R2), error reduction (lower MSE), and measurement stability enable reliable analysis of NO2- and NH4+ using common smartphones. Moreover, the modular design can be readily adapted for other analytes and incorporated into mobile or cloud-based analytical platforms for scalable environmental monitoring.

Livestream e-commerce now exceeds USD 500 billion in annual online retail sales. Nevertheless, a large proportion of viewers watch livestreams without sound, reducing engagement and lowering purchase conversion rates. In this work, we explore vision-based content understanding through lip reading and visual speech recognition (VSR) for automatic subtitle generation to enhance livestream clarity. We present LiveVSR-Net, a novel architecture combining 3D convolutional neural networks with transformer-based temporal modeling and hybrid CTC-attention decoding. Domain-adapted enhancements include e-commerce-specific vocabulary expansion, fine-tuned language model integration, and multi-speaker handling techniques. We conduct comprehensive ablation studies to quantify the contribution of each architectural component, evaluate on standard benchmarks (LRS2, LRS3, GRID), and provide real-world robustness analysis with deployability guidelines for livestreaming. Across three standard VSR benchmarks, the proposed LiveVSR-Net framework achieves competitive performance: 23.5% WER on LRS2, 25.1% WER on LRS3, and 0.9% on GRID. By combining 3D convolutional visual encoders, Conformer-based temporal modeling, hybrid CTC-attention decoding, and domain-adapted language models, the method achieves real-time recognition suitable for deployment in production environments.

The Financial Technology (FinTech) sector has witnessed rapid growth, resulting in increasingly complex and high-volume digital transactions. Although this expansion improves efficiency and accessibility, it also introduces significant vulnerabilities, including fraud, money laundering, and market manipulation. Traditional anomaly detection techniques often fail to capture the relational and dynamic characteristics of financial data. Graph Neural Networks (GNNs), capable of modeling intricate interdependencies among entities, have emerged as a powerful framework for detecting subtle and sophisticated anomalies. However, the high-dimensionality and inherent noise of FinTech datasets demand robust feature selection strategies to improve model scalability, performance, and interpretability. This paper presents a comprehensive survey of GNN-based approaches for anomaly detection in FinTech, with an emphasis on the synergistic role of feature selection. We examine the theoretical foundations of GNNs, review state-of-the-art feature selection techniques, analyze their integration with GNNs, and categorize prevalent anomaly types in FinTech applications. In addition, we discuss practical implementation challenges, highlight representative case studies, and propose future research directions to advance the field of graph-based anomaly detection in financial systems.

Visual speech recognition (VSR) aims to infer spoken content from visual observations of articulatory movements. Despite significant progress, it remains a challenging task in computer vision and speech processing. Its difficulty arises from pronounced speaker-to-speaker variability, the presence of homophenes (phonemes that are visually indistinguishable), changes in illumination, and the intrinsically high-dimensional nature of spatiotemporal lip dynamics. In this work, we propose NestLipGNN, a graph-based framework that integrates Graph Neural Networks (GNNs) with a nested multi-granularity learning strategy for visual speech recognition. We construct dynamic lip graphs from facial landmarks to model both spatial relationships between lip regions and their temporal motion during speech articulation. The proposed nested learning architecture supports hierarchical feature extraction across several levels of linguistic abstraction, spanning phoneme-level articulatory units, viseme-level visual speech categories, and word-level semantic representations. We further introduce a Temporal Graph Attention mechanism (T-GAT) that adaptively reweights the importance of distinct lip regions over time. We also introduce a graph-based contrastive learning objective to improve the discrimination of visually similar speech patterns, directly confronting the challenge of homophene resolution. Experiments on the LRW, LRS2, LRS3, and GRID datasets show that NestLipGNN improves recognition accuracy compared with existing methods, obtaining 92.3% word-level accuracy on LRW and delivering a 2.1% absolute performance gain over prior methods. Comprehensive ablation analyses confirm the contribution of each architectural component.

Neuromorphic processors that execute spiking neural network (SNN) inference through event-driven, spike-based computation achieve milliwatt-scale power consumption and sub-millisecond latency, properties that make them attractive for always-on edge intelligence. However, no systematic framework exists for determining when neuromorphic inference at the device level outperforms offloading to conventional edge or cloud accelerators within a collaborative pipeline. This paper addresses that gap by proposing a four-tier collaborative architecture that explicitly distinguishes the neuromorphic edge from the general-purpose edge and by deriving an energy-latency-accuracy offloading decision boundary between these tiers. We compare five neuromorphic platforms (Loihi 2, SpiNNaker 2, Akida, TrueNorth, DYNAP-SE) against three conventional edge processors using published benchmark data, revealing a 10--100$\times$ energy advantage for spiking inference on sparse workloads alongside a 3--10 percentage point accuracy gap that constrains deployment. Three application case studies, autonomous vehicle perception, industrial IoT predictive maintenance, and wearable health monitoring, ground the framework in concrete deployment scenarios and show how the offloading boundary changes across domains. We further analyze the simulation-to-hardware gap in SNN training as a key factor governing the accuracy constraint. The proposed framework provides system designers with a quantitative tool for placing neuromorphic inference within heterogeneous edge--cloud deployments.

Pill image recognition is critical for medication safety, yet existing systems struggle with the effective integration of visual and textual information from pill imprints. This paper presents a comprehensive comparative study of six state-of-the- art Optical Character Recognition (OCR) methods for imprint extraction, including SAR, PaddleOCR, Tesseract, CRNN, RobustScanner, and TrOCR, and investigates the application of Vision Transformer (ViT) architecture for multimodal feature fusion. We systematically evaluate six OCR methods—Tesseract (traditional), CRNN (CNN-RNN), SAR (attention-based), RobustScanner (hybrid attention), PaddleOCR (hybrid architecture), TrOCR (transformer-based)—on the CURE pill dataset using identical experimental protocols to ensure fair comparison despite the dataset’s challenging uncontrolled imaging conditions. RobustScanner achieves highest standalone accuracy (95.80% test accuracy, 92.35% mAP), while PaddleOCR provides optimal accuracy-speed balance (95.66% test accuracy, 92.02% mAP, 3.06 FPS). For multimodal fusion, Vision Transformer significantly outperforms traditional approaches: when combined with PaddleOCR, ViT fusion achieves 95.66% test accuracy and 92.02% test mAP, representing substantial improvements of +3.34pp accuracy and +5.69pp mAP over Simple Sum fusion (92.32% accuracy, 86.33% mAP), and +0.86pp accuracy and +1.37pp mAP over Weighted Sum fusion (94.80% accuracy, 90.65% mAP). Averaging across all six OCR methods with ViT fusion, the framework demonstrates mean test accuracy of 95.18% at 2.30 FPS, validating practical viability for real-world deployment. This study provides the first systematic comparison of modern OCR architectures specifically for pharmaceutical imprint extraction.

Stock price prediction is the most enduring problem of financial engineering because financial markets are nonlinear, non-stationary, and informationally dense. Classical models based on historical price trends and fundamentals often miss the context of qualitative financial narratives. This paper proposes FinLLM-Predict, a new multimodal deep learning framework based on Large Language Models (LLMs) and quantitative financial signals for directional price forecasting and volatility estimation. In our architecture, we combine domain-specific LLMs tuned to semantic analysis of financial news, earnings transcripts, and technical indicators based on price series via a market-responsive dynamic cross-modal attention mechanism adjusting modality contributions. Hierarchical multitask learning predicts price direction, magnitude, and realized volatility simultaneously for coordinated optimization on related objectives. Experiments conducted over a six-year horizon on the S&P 500 index show that FinLLM-Predict achieves a 12.3% improvement in directional accuracy and an 18.7% reduction in volatility prediction error over state-of-the-art baselines, and is particularly robust in high-volatility regimes. Ablation studies show that each component is needed, and attention analyses indicate that textual signals become more informative during periods of market uncertainty, consistent with economic intuition. The interpretability, performance, and adaptability of the framework make it an interesting instrument for algorithmic trading and risk management.

Vietnamese visual speech recognition remains challenging because public resources are limited and many phrases exhibit similar mouth movements over short temporal spans. This paper investigates a Vietnamese phrase-level dataset using a visual-only architecture that combines Hierarchical Deep Aggregation (HDA), CenterNet-based lip localization, and a Spatio-Temporal Graph Neural Network (ST-GNN). The experimental subset is constructed from $\text{5, 1 3 0}$ labeled clips covering $\text{7 2}$ common Vietnamese phrases and is reduced to 512 validated videos spanning 11 phrases through timestamp-based cropping, face-mesh filtering, and phrase-frequency selection. The final dataset is divided into 80% training, 10% validation, and 10% test sets. Experimental results show that the full HDA-STGNN model achieves the best performance, reaching 35.7% WER and 89.5% accuracy, outperforming HDA only and the partial variants with CenterNet or ST-GNN alone. These findings indicate that accurate lip localization and graph-based temporal modeling are both beneficial for Vietnamese phrase-level visual speech recognition on a small but controlled dataset.

Financial fraud is threatening global economic stability, with annual losses exceeding USD 5 trillion. Traditional batch-processing detection systems, limited by inherent latency, are inadequate to defend against sophisticated real-time attacks. This paper introduces StreamGuard, a novel AI-driven framework for real-time fraud detection that leverages streaming data analytics and an adaptive ensemble architecture. StreamGuard aggregates Long Short-Term Memory (LSTM) networks, Gradient Boosting Machines (GBM) and Isolation Forests into a dynamic ensemble governed by real-time performance metrics to suppress concept drift. It adopts a multi-scale sliding window feature engineering pipeline, incremental model updating via online learning and an explainable AI module based on SHAP values for regulatory compliance. StreamGuard achieves 97.3% detection accuracy with average latency of 12.3 milliseconds, outperforming state-of-the-art methods in AUC-ROC and reducing the false positive rate by 34% on three real-world financial datasets. The framework supports throughput greater than 100,000 transactions per second, keeping pace with modern financial infrastructure requirements while remaining interpretable and adaptable.

Ensuring the reliability of power transmission networks depends heavily on the early detection of faults in key components such as insulators, which serve both mechanical and electrical functions. Even a single defective insulator can lead to equipment breakdown, costly service interruptions, and increased maintenance demands. While unmanned aerial vehicles (UAVs) enable rapid and cost-effective collection of high-resolution imagery, accurate defect identification remains challenging due to cluttered backgrounds, variable lighting, and the diverse appearance of faults. To address these issues, we introduce a real-time inspection framework that integrates an enhanced YOLOv10 detector with a Hybrid Quantum-Enhanced Graph Neural Network (HQGNN). The YOLOv10 module, fine-tuned on domain-specific UAV datasets, improves detection precision, while the HQGNN ensures multi-object tracking and temporal consistency across video frames. This synergy enables reliable and efficient identification of faulty insulators under complex environmental conditions. Experimental results show that the proposed YOLOv10-HQGNN model surpasses existing methods across all metrics, achieving Recall of 0.85 and Average Precision (AP) of 0.83, with clear gains in both accuracy and throughput. These advancements support automated, proactive maintenance strategies that minimize downtime and contribute to a safer, smarter energy infrastructure.

Vietnamese traditional medicine (VTM) relies on a rich diversity of native plant species; however, urbanization and the dominance of Western medicine have made many of them increasingly difficult to locate and conserve. Accurate identification of these medicinal plants is therefore critical for preserving VTM knowledge and integrating it into digital healthcare and educational platforms. This study introduces DualBranchNet, a deep learning classifier designed to recognize Vietnamese medicinal plants in the VNPlant200 image dataset. DualBranchNet employs a dual-branch convolutional architecture that concatenates features extracted by two pretrained backbones, DenseNet-201 and Xception. To improve training stability and generalization, we employ learning rate warm-up, L2 regularization, and on-the-fly data augmentation. Model performance is tracked via Top-1 accuracy and categorical cross-entropy, and the best checkpoint is evaluated on the full validation set. Experimental results show that DualBranchNet attains high classification accuracy and represents a reliable step toward digitizing and preserving Vietnamese medicinal plant knowledge.

Visual Speech Recognition (VSR) for Vietnamese has not been extensively explored because of the lack of VSR datasets for the Vietnamese language. In this study, we propose the first Vietnamese Lip Reading (VLR) corpus, namely Vietnamese Lip Reading, consisting of 198 hours of speech and lip data captured from 1,781 videos extracted from online video platforms. The proposed corpus contains word-level alignments with sufficiently high-resolution mouth frames and rich head pose variations with several lighting setups and occlusions. Experiments using six VSR networks based on deep learning approach on VLR achieved an average word error rate of 62.18% when using AV-HuBERT as the base model. To improve on the performance of VSR, a refinement pipeline using a combination of shallow fusion of a Vietnamese 5-gram language model with a difficulty filtering mechanism via curriculum learning, region-of-interest (ROI) detection for lip alignment, and re-ranking using N-best with a lightweight Transformer-based decoder is proposed. Results showed that the proposed refinements were able to bring down the average word error rate to about 30%. The proposed Vietnamese Lip Reading corpus serves as a VSR benchmark dataset enables a systematic exploration of several aspects of VSR on the Vietnamese language, as well as a starting point to develop VSR for other low-resource and tonal languages.

Fine-grained image classification remains challenging because subtle inter-class differences and large intra-class variations limit existing supervised contrastive learning methods. Although sigmoid-based methods such as SCS-SupCon alleviate negative-sample dilution through adaptive pairwise decision boundaries, they still suffer from batch-limited negative-sample sparsity and unstable boundaries caused by the lack of historical feature references. We propose B-SupCon (Sigmoid Supervised Contrastive Learning with Memory Bank for Feature Disentanglement), which augments sigmoid-based supervised contrastive learning with recent historical embeddings. Specifically, B-SupCon introduces an Adjustable Memory Bank to enlarge current/history pairwise comparisons and an adaptive variant with a learnable utilization ratio ρ to control the amount of historical information used, while retaining the style-distance constraint of SCS-SupCon to better separate common and style-related features. Experiments on five benchmarks, including CUB200-2011 and Stanford Dogs, show that B-SupCon and B-SupCon (Adaptive) consistently outperform SCS-SupCon and strong InfoNCE-based supervised contrastive baselines across CNN and Transformer backbones, while remaining competitive with recent clustering-based and hybrid methods. These results confirm the effectiveness of the proposed Memory Bank mechanisms and style regularization.

Smartphone-based colorimetric sensing is a promising low-cost alternative for rapid chemical analysis, but its robustness remains influenced by image preprocessing, device-dependent color variation, and the choice of learning architecture. In this work, we study a compact classification-first problem: binary analyte identification of ammonium $\left( {{\text{NH}}_4^ + } \right)$ versus nitrite $\left( {{\text{NO}}_2^ - } \right)$ from cropped smartphone images of colorimetric test strips. Using a laboratory benchmark of approximately 10K strip images, we evaluate an image-native deep pipeline consisting of grouped data splitting, YOLO-based region-of-interest extraction, optional green-border mean normalization, and deep classification. The study is intentionally restricted to a single explicit calibration comparison, none versus green(mean), and five representative backbones: MobileNetV3-Large, EfficientNetB0, TF-EfficientNet-B3, ConvNeXt-Tiny, and Swin-Tiny. Results show that deep classification on cropped strip images is already strong, but the effect of explicit green-border normalization is heterogeneous rather than uniformly beneficial: it improves some backbones, changes others only marginally, and degrades several settings. The strongest overall result among the main models is obtained without explicit calibration. These findings suggest that, relative to earlier feature-based smartphone colorimetry, deep image-based classification reduces dependence on explicit calibration, but does not eliminate it; the impact remains backbone-dependent. The proposed benchmark provides a concise and deployment-relevant reference for future smartphone colorimetry studies focused on learned visual classification.