Video anomaly detection identifies unusual behaviors or objects in video streams, crucial for public safety. Traditional methods often fail to capture subtle motion cues from raw frames, while recent deep learning models-though more effective-require heavy computational resources, limiting their real-time applicability. To address these challenges, we propose an unsupervised, lightweight, and efficient model: Flow-Enhanced Anomaly Detection (FEAD). Our key contribution is a novel optical flow-guided feature fusion mechanism that leverages pre-extracted optical flow information as a dynamic prior, effectively directing the raw image stream to focus on critical motion regions during feature extraction. By combining raw video frames and optical flow maps in a dual-stream architecture, FEAD predicts future frames and detects anomalies via reconstruction error. This approach enhances sensitivity to motion irregularities while maintaining low computational cost, making it suitable for real-time deployment in resource-constrained environments. We evaluated FEAD on three benchmark datasets-Ped2, Avenue, and ShanghaiTech-achieving AUC scores of 98.4%, 87.1%, and 75.6% respectively. In addition to strong accuracy, FEAD shows substantial advantages in inference speed and efficiency compared to existing methods, underscoring its practical value for real-world applications in surveillance and public safety.

Abstract In this work, we evaluate the Kaczmarz, Coordinate Descent, and
Cimmino algorithms together with the resolution principle as stopping criteria,
using a synthetic signal model for the MAST-U FILD, complementing the
efforts recently done for the ASDEX Upgrade FILD. The performance of these
algorithms is assessed by analyzing the evolution of the reconstruction error
as well as the computation time. To further assess the reliability of the
reconstructions, a ”fidelity map” that reconstructs signals at each grid point
is introduced to visualize reconstruction accuracy across velocity space. The
Kaczmarz algorithm, which shows better performance in terms of accuracy, is
also applied to experimental MAST-U FILD measurements of prompt fast-ion
losses in an L-mode plasma heated by an on-axis Neutral Beam Injector (NBI)
with 1.5 MW of input power. This algorithm demonstrates improved performance
compared to 0th-order Tikhonov regularization

Abstract Existing methods predominantly perform denoising and matching within a unified visual feature space, often neglecting multi-view positional and color information. This paper proposes a method based on noise loop detection and visual complementary fusion for high-precision Three-Dimensional(3D) reconstruction. Specifically, leveraging the association between the robotic arm coordinate system and the camera coordinate system, we first introduce a Multi-level Loop Detection Noise(MLDN) module. This module identifies and processes noise in the data in stages, mitigating its adverse effects on reconstruction accuracy. Subsequently, spatial transformations are applied to both RGB and depth images, establishing a robust initial foundation for subsequent processing. Finally, leveraging the complementary characteristics of both image types, we design an RGB and Depth image (RGB-D) Collaborative Registration (RDCR) module to maximize the strengths of each visual modality. Experiments on four public datasets and real-world components demonstrated a 1.26mm reduction in contour reconstruction error, a 1.4% improvement in IoU, a -1.07mm decrease in Ch-d, and a 5.3% increase in CR compared to state-of-the-art methods. The proposed framework offers a novel approach for high-precision 3D visual reconstruction.

Purpose This study aim to propose a low-power, edge-deployable out-of-distribution (OOD) anomaly detection model and an anomaly-triggered time synchronization mechanism to address the challenges of anomaly detection accuracy and spatiotemporal consistency in Industrial Internet of Things (IIoT) systems, particularly for resource-constrained TinyML platforms. Design/methodology/approach The authors developed a hybrid anomaly detection model based on a one-dimensional convolutional autoencoder (1D-CAE) and one-class support vector machine. The model learns normal patterns through reconstruction error and performs anomaly detection. In addition, the authors introduce a lightweight anomaly-driven synchronization mechanism that combines regression-based clock drift correction with adaptive heartbeat signaling to ensure clock synchronization between nodes. Findings In the experimental environment we built, the authors’ proposed method achieved an anomaly detection accuracy of 99.8% with a timestamp alignment precision of 0.8 ms, while maintaining low power and bandwidth consumption. Compared to traditional methods, this approach reduces bandwidth consumption by 79.2% and energy consumption by 5.5%, demonstrating its efficiency and feasibility for real-world IIoT deployments. Originality/value The authors integrate OOD anomaly detection with an anomaly-triggered synchronization mechanism into a unified framework, specifically designed for resource-constrained TinyML devices. This provides an innovative solution for achieving spatiotemporal consistency in edge-based anomaly analytics. This approach holds significant theoretical and practical value for advancing efficient anomaly detection and real-time synchronization in industrial IoT applications.

ABSTRACT In open-set hyperspectral image (HSI) classification, existing methods commonly depend on reconstruction error to detect unknown classes. However, commonly used residual or dense connections are often unable to capture both local and global features effectively, and a single linear layer cannot adequately express the complex structure of nonlinear data. These limitations restrict both classification accuracy and the effectiveness of recognizing unknow classes. To address these shortcomings, we propose a compact classification strategy based on class anchors, which incorporates both global and local feature extraction methods and improves the representation of complex structures by replacing the linear layer with a Kolmogorov-Arnold Networks Linear (KANLinear) layer. Based on this approach, we propose a dual-branch open-set HSI classification method integrating KANLinear and class anchor clustering (DOHKC). First, we use a dual-branch module with residual and dense connections to extract spectral and spatial features. The extracted features then pass through a global average pooling layer followed by a KANLinear layer to generate a logit vector Z . Finally, classification is conducted using the class anchor clustering module. Our method is validated on three HSI datasets, demonstrating not only improved classification accuracy for known classes but also enhanced recognition of unknown classes.

Extracellular vesicles (EVs) are increasingly recognized as promising non-invasive biomarkers for cancer and other diseases, but their clinical translation remains limited by the lack of comprehensive characterization strategies. Spectroscopic approaches such as Fourier-transform infrared (FTIR) spectroscopy can provide a global biochemical fingerprint of intact EVs, but their interpretation requires advanced analytical tools. In this study, we applied an autoencoder-based framework to attenuated total reflection FTIR (ATR-FTIR) spectra of blood-derived components, including plasma, red blood cells (RBCs), RBC-ghosts, and EVs, comprising 278 samples collected from 135 patients, to obtain latent features capable of capturing biologically meaningful variability. The autoencoder compressed spectra into 12 latent features while preserving spectral information with low reconstruction error. Unsupervised UMAP projection of the latent features separated the blood components into different clusters, supporting their biological relevance. The model was then applied to EV spectra from patients with hepatocellular carcinoma (HCC) and cirrhotic controls. Four features significantly differed between the two groups, and an elastic-net regularized logistic model evaluated with a leave-one-out cross-validation framework retained a single latent feature, achieving an out-of-fold ROC AUC of 0.785 (95% CI 0.602–0.967), with performance broadly comparable to that typically reported for AFP, the most commonly used biomarker for HCC. This study provides the first proof-of-concept that an autoencoder can be applied to FTIR spectra of EVs, extracting biologically relevant latent features with potential application in cancer detection.

Spatiotemporal monitoring of tissue coagulation treated by continuous high-intensity focused ultrasound with pulse-inversion shear wave elastography.

Post-training quantization (PTQ) for transformer-based large foundation models (LFMs) significantly accelerates model inference and relieves memory constraints, without incurring model training. However, existing methods face three main issues: 1) The scaling factors, which are commonly used in scale reparameterization based weight-activation quantization for mitigating the quantization errors, are mostly hand-crafted defined which may lead to suboptimal results; 2) The formulation of current quantization error defined by L2-norm ignores the directional shifts after quantization; 3) Most methods are devised tailored for single scenario, i.e., only evaluated on LLMs or only designed for weight-only quantization, which lacks of a comprehensive evaluation on diverse benchmarks and a broad application scope. To address these challenges, this paper introduces a unified Learnable and Robust post-training Quantization framework for transformer based LFMs and various quantization scenarios, called LRQuant. First, we consider an efficient block-wise learnable paradigm to find optimal scaling factors which are initialized by logarithmic activation equivalent and get suitable clipping range of quantization steps. In addition, we empirically find that only relying on MSE loss could hardly lead to optimal quantization results, so we reformulate the quantization error and then propose a novel loss function based on the negative logarithm of cosine similarity (NLC loss) between outputs of full-precision and quantized block. To fully investigate the potentiality of our learnable paradigm, we propose a more superior version LRQuant+. Specifically, we first propose a dynamically weighted scheme to balance MSE and NLC loss, and then devise learnable rotation vectors to further directly reduce directional gaps. In addition, we improve the block-wise optimization framework into a novel two-branch nature which jointly considers the error propagation and homologous reconstruction error. Extensive experiments demonstrate the superiority of our LRQuantand LRQuant+, as well as their unified effectiveness across various LFMs for both weight-activation and weight-only quantization, especially under challenging quantization scenarios, i.e., W4A4 and W2A16 on LLMs, ViTS, and MLLMs.

.SignificanceCurrent optical health-sensing devices rely on simplified homogeneous tissue models or semi-empirical ratiometric methods, which inadequately address anatomical complexity and inter-individual optical variability. This introduces systematic errors in light propagation modeling, compromising measurement accuracy and clinical robustness, necessitating organ-specific optical models for reliable physiological sensing.AimTo develop a standard optical digital wrist (DW) model by integrating magnetic resonance imaging (MRI) and diffuse optical imaging (DOI), enabling anatomically accurate and optically realistic modeling of wrist tissues for improved precision in wearable optical health monitoring applications.ApproachThe multimodal MRI-DOI framework was implemented, comprising three key components: (1) statistical integration of high-resolution MRI datasets generated a population-averaged anatomical DW template; (2) region-based time-domain diffuse optical tomography (TD-DOT) with MRI-derived anatomical priors, extracted depth-resolved optical properties of subsurface tissues; (3) spatial frequency domain imaging (SFDI) supplemented high-resolution optical properties of superficial skin layers.ResultsSimulation experiments demonstrated the high accuracy of region-based TD-DOT reconstruction, with mean errors below 8.57% () and 9.63% (), quantitatively supporting the precision of the proposed approach. Phantom experiments with wrist-mimicking phantoms yielded mean reconstruction errors of 10.52% () and 13.23% () for TD-DOT, and the SFDI top-layer quantification yielded lower errors of 4.48% () and 8.69% (), validating the performance of the TD-DOT system and the SFDI system. Furthermore, in vivo optical property measurements showed strong agreement with literature values, further validating the reliability and practicality of the methodology.ConclusionsWe establish a standard DW template and develop an in vivo optical structure acquisition methodology, transitioning biosensing models from homogeneous approximations to anatomically layered models. The approach can enhance the customization, dynamic adaptability, and clinical validity of biosensing technologies.

Adaptive regularization weight selection for compressed sensing MRI reconstruction.

Background: With the rapid evolution of single-cell RNA sequencing technology, the study of cellular heterogeneity in complex tissues has reached an unprecedented resolution. One critical task of the technology is cell-type annotation. However, challenges persist, particularly in annotating novel cell types. Objective: Current methods rely heavily on well-annotated reference data, using correlation comparisons to determine cell types. However, identifying novel cells remains unstable due to the inherent complexity and heterogeneity of scRNA-seq data and cell types. To address this problem, we propose scADCA, a method based on anomaly detection, for identifying novel cell types and annotating the entire dataset. Methods: The convolutional modules and fully connected networks are integrated into an autoencoder, and the reference dataset is trained to obtain the reconstruction errors. The threshold based on these errors can distinguish between novel and known cells in the query dataset. After novel cells are identified, a multinomial logistic regression model fully annotates the dataset. Results: Using a simulation dataset, three real scRNA-seq pancreatic datasets, and a real scRNA-seq lung cancer cell line dataset, we compare scADCA with six other cell-type annotation methods, demonstrating competitive performance in terms of distinguished accuracy, full accuracy, F!-score, and confusion matrix. Conclusion: In conclusion, the scADCA method can be further improved and expanded to achieve better performance and application effects in cell type annotation, which is helpful to improve the accuracy and reliability of cytology research and promote the development of single-cell omics.

The paper addresses the topic of detecting money laundering operations in transaction data represented as graph data-structures. We propose the integration of autoencoder components in Graph Neural Networks (GNN) architectures, in order to incorporate a reconstruction step in the traditional edge classification problem and enhance model quality based upon the usage of reconstruction errors. We show that enhancing GNNs with autoencoder components improves the predictive performance of money laundering detection, on data represented as homogeneous graphs. Additionally, the Shapley value is computed in order to gain further insight into the most important features from distinguishing normal and fraudulent activities.

Abstract With the expansion of the power system scale and the increasing complexity of the grid structure, the probability of faults occurring during transmission line operation continues to increase. Especially in multi-branch topology structures, the diversity of fault wave propagation paths poses severe challenges to traditional ranging methods. Therefore, this study proposes a traveling wave fault location method for transmission lines that combines improved SAA and VMD. This method optimizes the sparrow search process to adjust the key parameters of VMD, and combines the branch structure of the line with the principle of dual end distance measurement to construct a segmented recognition strategy, achieving accurate distance measurement in multi branch environments. In performance testing, when the SNR was 30 dB, the proposed method improved the modal separation and reconstruction error by approximately 35.85% and 41.86% compared to the traditional SAA. In application testing, the proposed method showed errors of 91.6 m, 96.8 m, 95.4 m, and 99.7 m at each distance measurement point in a two-phase grounding fault scenario. The experiment showed that this method exhibited good ranging accuracy and stability under various typical working conditions, and had certain practicality and promotion value. This study provides more real-time and engineering adaptive technical support for the monitoring of transmission line status and rapid fault response.

This study investigates three types of emission families in Hidden Markov Models (HMMs) for reconstructing Bedoyo Majapahit dance motion captured using a markerless system. The recorded skeleton data, consisting of 3,341 frames and 33 joints per frame, were normalized and reduced into a 30-dimensional latent space using Principal Component Analysis (PCA). Three emission variants were evaluated: single-Gaussian HMM, Gaussian-mixture HMM (GMM-HMM), and Multinomial HMM. The evaluation employed a tri-metric scheme consisting of Mean Squared Error (MSE), Dynamic Time Warping (DTW), and Fréchet distance to measure reconstruction fidelity. The experimental results showed that GMM-HMM consistently outperformed the other two models, achieving the lowest reconstruction error and the closest alignment to the original temporal and geometric motion profiles. The Gaussian HMM demonstrated moderate performance but tended to underestimate motion amplitude, while the Multinomial HMM produced the weakest results due to the discretization of continuous pose data. These findings indicate that multimodal emission functions provide a more expressive representation for continuous dance motion. The study highlights the suitability of GMM-HMM for traditional dance preservation through computational modeling and contributes to the development of digital motion archiving for cultural heritage.

Abstract Anomaly detection in high-energy physics is essential for identifying new physics beyond the Standard Model. Autoencoders provide a signal-agnostic approach but are limited by the topology of their latent space. This work explores topology-aware autoencoders, embedding phase-space distributions onto compact manifolds that reflect energy-momentum conservation. We construct autoencoders with spherical ( S n ), product ( S 2 ⊗ S 2 ), and projective ( RP 2 ) latent spaces and compare their anomaly detection performance against conventional Euclidean embeddings on simulated collider events. Our results show that autoencoders with topological priors improve anomaly separation by preserving the global structure of the data manifold and reducing spurious reconstruction errors. Applying our approach to simulated hadronic top-quark decays, we show that latent spaces with appropriate topological constraints enhance sensitivity and robustness in detecting anomalous events. This study establishes topology-aware autoencoders as a powerful tool for unsupervised searches for new physics in particle-collision data.

Video Anomaly Detection is a critical task for identifying unusual events in video streams, with applications ranging from public safety surveillance to industrial monitoring. Traditional VAD methods, often based on reconstruction or prediction errors, excel at detecting deviations but typically lack semantic understanding, failing to explain why an event is anomalous. The recent advent of Vision-Language Models and Large Language Models has introduced a new paradigm, enabling systems to interpret and reason about video content in natural language. However, existing VLM/LLM-based approaches often focus either on rich, open-ended description or on structured, rule-based reasoning, but rarely both. In this paper, we address this gap by proposing a novel hybrid framework that synergizes the strengths of descriptive and deductive models. Our approach first leverages a powerful VLM to generate detailed, contextual scene descriptions. These descriptions are then fed into a rule-driven LLM, which uses a pre-induced set of contextual rules to make a final anomaly judgment and provide a human-readable explanation grounded in the specific rule that was violated. We validate our approach on the large-scale UCF-Crime dataset and conduct an analysis of key hyperparameters, including the VLM's input prompt and the number of frames used for description. Our results demonstrate the effectiveness of the proposed architecture and offer insights into building more interpretable, reliable, and context-aware VAD systems.

Background: The asymmetric distribution of optical photons near the edges of Positron Emission Tomography (PET) sensor modules introduces errors in the coordinate reconstruction of scintillation events when center-of-gravity (CoG) algorithms are utilized. This issue, sometimes referred to as the “edge effect”, results in overlap of crystal pixel signatures in flood maps and potential image artifacts in reconstructed PET images. Methods: Partly segmented 5 mm thick borosilicate light guides with slits cut parallel to the edges are filled with barium sulfate to restrict the spread of optical photons near the edges of the light guide. Data acquisitions are performed using single PET sensor modules in coincidence, both with single and multiplexed channel readout. CoG and truncated center-of-gravity (TCoG) methods are used for coordinate reconstruction. Results: A 22 × 22 array of crystal signatures are distinguishable on crystal flood maps produced using sensor modules with solid light guides and 24 × 24 arrays can be identified when using a partly segmented light guide. The pixel resolution around the edges and corners of the flood map is further improved when TCoG is used for coordinate reconstruction. Summary: We show that the introduction of a partly segmented light guide greatly improves coordinate reconstruction accuracy at the edges of a sensor module. In addition, it is demonstrated that the partly segmented light guides can be used in parallel with other proposed methods designed to fix the “edge effect”, including TCoG, to further coordinate reconstruction improve accuracy and crystal flood map quality.

Abstract The safe operation of process industries requires accurate anomaly monitoring models that can effectively model intricate time-varying and nonlinear variable dynamics to minimize false alarms. This paper introduces the Gated Recurrent Deep Shrinkage Variational Autoencoder (GDVAE), a novel framework for dynamic nonlinear process monitoring. The method first utilizes a recursive Variational Autoencoder (VAE) to capture temporal dependencies and characterize the data distribution from normal operations, enabling precise online data reconstruction. Following this, a deep residual shrinkage network (DRSN) is employed to assign unique weights to process variables based on their fault relevance within a given time window. Finally, a novel monitoring statistic is designed by synergizing the reconstruction error and the assigned fault feature weights for real-time process monitoring. We validated our method on the benchmark Tennessee Eastman (TE) process. Comparative analyses demonstrate that the proposed GDVAE model achieves significant improvements in detection accuracy and reliability over conventional Dynamic Principal Component Analysis (DPCA), Kernel Principal Component Analysis (KPCA), and standard VAE methods, offering a more robust solution for dynamic nonlinear process monitoring.

Insider threats pose severe risks to organizations, as authorized users can misuse legitimate access to cause damage. This paper presents an AI-powered insider threat detection framework that models user behavior through endpoint activity data. The system employs a Bidirectional LSTM Autoencoder to learn normal behavioral patterns and detect anomalies via reconstruction error, enhanced by an Isolation Forest for reducing false positives. A multi-factor threat scoring engine evaluates anomaly intensity, frequency, and recency to assess user risk levels. Experimental results on simulated enterprise data achieved 93.2% accuracy with a 5.1% false positive rate, demonstrating effective behavioral anomaly detection and real-time risk visualization through a Streamlit dashboard.

Abstract Automated and non‐invasive anomaly detection methods are critical for ensuring operational safety and continuity on intelligent construction sites. This study proposes a novel unsupervised audio signal processing framework for real‐time monitoring of construction equipment based on their operational acoustic signatures. The proposed method relies exclusively on historical data from normal operations to characterize temporal audio patterns, enabling the detection of previously unseen anomalies without requiring labeled anomaly data for training. It extracts 39 acoustic features from raw waveform audio and reconstructs them using a temporal convolutional network autoencoder. Anomalies are identified by monitoring the reconstruction errors through a multivariate cumulative sum (MCUSUM) statistical process control chart. Upon detecting an anomaly, the method identifies contributing acoustic features via correlation maximization decomposition of MCUSUM statistics. The proposed method detected 100% of anomalies in 50 real‐world slider rail tests, with an average detection time of 2.15 s post onset.