Search Results (6)

Micro-expressions, characterized by brief and subtle facial muscle movements, are essential for conveying nuanced emotions in digital humans, yet existing rendering techniques often produce rigid or emotionally monotonous animations due to the inadequate modeling of temporal dynamics and action unit interdependencies. This paper proposes a graph-driven framework for micro-expression rendering that generates emotionally diverse and lifelike expressions. We employ a 3D-ResNet-18 backbone network to perform joint spatio-temporal feature extraction from facial video sequences, enhancing sensitivity to transient motion cues. Action units (AUs) are modeled as nodes in a symmetric graph, with edge weights derived from empirical co-occurrence probabilities and processed via a graph convolutional network to capture structural dependencies and symmetric interactions. This symmetry is justified by the inherent bilateral nature of human facial anatomy, where AU relationships are based on co-occurrence and facial anatomy analysis (as per the FACS), which are typically undirected and symmetric. Human faces are symmetric, and such relationships align with the design of classic spectral GCNs for undirected graphs, assuming that adjacency matrices are symmetric to model non-directional co-occurrences effectively. Predicted AU activations and timestamps are interpolated into continuous motion curves using B-spline functions and mapped to skeletal controls within a real-time animation pipeline (Unreal Engine). Experiments on the CASME II dataset demonstrate superior performance, achieving an F1-score of 77.93% and an accuracy of 84.80% (k-fold cross-validation, k = 5), outperforming baselines in temporal segmentation. Subjective evaluations confirm that the rendered digital human exhibits improvements in perceptual clarity, naturalness, and realism. This approach bridges micro-expression recognition and high-fidelity facial animation, enabling more expressive virtual interactions through curve extraction from AU values and timestamps. Full article

Industrial equipment fault diagnosis faces dual challenges: significant data distribution discrepancies caused by diverse operating conditions impair generalization capabilities, while underutilized spatio-temporal information from multi-source data hinders feature extraction. To address this, we propose a spatio-temporal collaborative perception-driven feature graph construction and topology mining methodology for variable-condition diagnosis. First, leveraging the operational condition invariance and cross-condition consistency of fault features, we construct fault feature graphs using single-source data and similarity clustering, validating topological similarity and representational consistency under varying conditions. Second, we reveal spatio-temporal correlations within multi-source feature topologies. By embedding multi-source spatio-temporal information into fault feature graphs via spatio-temporal collaborative perception, we establish high-dimensional spatio-temporal feature topology graphs based on spectral similarity, extending generalized feature representations into the spatio-temporal domain. Finally, we develop a graph residual convolutional network to mine topological information from multi-source spatio-temporal features under complex operating conditions. Experiments on variable/multi-condition datasets demonstrate the following: feature graphs seamlessly integrate multi-source information with operational variations; the methodology precisely captures spatio-temporal delays induced by vibrational direction/path discrepancies; and the proposed model maintains both high diagnostic accuracy and strong generalization capacity under complex operating conditions, delivering a highly reliable framework for rotating machinery fault diagnosis. Full article

This study presents a novel framework that integrates Vision Graph Neural Networks (ViGs) with supervised contrastive learning for enhanced spectro-temporal image analysis of speech signals in Parkinson’s disease (PD) detection. The approach introduces a frequency band decomposition strategy that transforms raw audio into three complementary spectral representations, capturing distinct PD-specific characteristics across low-frequency (0–2 kHz), mid-frequency (2–6 kHz), and high-frequency (6 kHz+) bands. The framework processes mel multi-band spectro-temporal representations through a ViG architecture that models complex graph-based relationships between spectral and temporal components, trained using a supervised contrastive objective that learns discriminative representations distinguishing PD-affected from healthy speech patterns. Comprehensive experimental validation on multi-institutional datasets from Italy, Colombia, and Spain demonstrates that the proposed ViG-contrastive framework achieves superior classification performance, with the ViG-M-GELU architecture achieving 91.78% test accuracy. The integration of graph neural networks with contrastive learning enables effective learning from limited labeled data while capturing complex spectro-temporal relationships that traditional Convolution Neural Network (CNN) approaches miss, representing a promising direction for developing more accurate and clinically viable speech-based diagnostic tools for PD. Full article

Knowledge graph embedding has been identified as an effective method for node-level classification tasks in directed graphs, the objective of which is to ensure that nodes of different categories are embedded as far apart as possible in the feature space. The directed graph is a general representation of unstructured knowledge graphs. However, existing methods lack the ability to simultaneously approximate high-order filters and globally pay attention to the task-related connectivity between distant nodes for directed graphs. To address this limitation, a directed spectral graph transformer (DSGT), a hybrid architecture model, is constructed by integrating the graph transformer and directed spectral graph convolution networks. The graph transformer leverages multi-head attention mechanisms to capture the global connectivity of the feature graph from different perspectives in the spatial domain, which bridges the gap between frequency responses and, further, naturally couples the graph transformer and directed graph convolutional neural networks (GCNs). In addition to the inherent hard inductive bias of DSGT, we introduce directed node positional and structure-aware edge embedding to provide topological prior knowledge. Extensive experiments demonstrate that the DSGT exhibits state-of-the-art (SOTA) or competitive node-level representation capabilities across datasets of varying attributes and scales. Furthermore, the experimental results indicate that the homophily and degree of correlation of the nodes significantly influence the classification performance of the model. This finding opens significant avenues for future research. Full article

Detecting emotions is a growing field aiming to comprehend and interpret human emotions from various data sources, including text, voice, and physiological signals. Electroencephalogram (EEG) is a unique and promising approach among these sources. EEG is a non-invasive monitoring technique that records the brain’s electrical activity through electrodes placed on the scalp’s surface. It is used in clinical and research contexts to explore how the human brain responds to emotions and cognitive stimuli. Recently, its use has gained interest in real-time emotion detection, offering a direct approach independent of facial expressions or voice. This is particularly useful in resource-limited scenarios, such as brain–computer interfaces supporting mental health. The objective of this work is to evaluate the classification of emotions (positive, negative, and neutral) in EEG signals using machine learning and deep learning, focusing on Graph Convolutional Neural Networks (GCNN), based on the analysis of critical attributes of the EEG signal (Differential Entropy (DE), Power Spectral Density (PSD), Differential Asymmetry (DASM), Rational Asymmetry (RASM), Asymmetry (ASM), Differential Causality (DCAU)). The electroencephalography dataset used in the research was the public SEED dataset (SJTU Emotion EEG Dataset), obtained through auditory and visual stimuli in segments from Chinese emotional movies. The experiment employed to evaluate the model results was “subject-dependent”. In this method, the Deep Neural Network (DNN) achieved an accuracy of 86.08%, surpassing SVM, albeit with significant processing time due to the optimization characteristics inherent to the algorithm. The GCNN algorithm achieved an average accuracy of 89.97% in the subject-dependent experiment. This work contributes to emotion detection in EEG, emphasizing the effectiveness of different models and underscoring the importance of selecting appropriate features and the ethical use of these technologies in practical applications. The GCNN emerges as the most promising methodology for future research. Full article

Transfer learning is a modern concept that focuses on the application of ideas, models, and algorithms, developed in one applied area, for solving a similar problem in another area. In this paper, we identify links between methodologies in two fields: video prediction and spatiotemporal traffic forecasting. The similarities of the video stream and citywide traffic data structures are discovered and analogues between historical development and modern states of the methodologies are presented and discussed. The idea of transferring video prediction models to the urban traffic forecasting domain is validated using a large real-world traffic data set. The list of transferred techniques includes spatial filtering by predefined kernels in combination with time series models and spectral graph convolutional artificial neural networks. The obtained models’ forecasting performance is compared to the baseline traffic forecasting models: non-spatial time series models and spatially regularized vector autoregression models. We conclude that the application of video prediction models and algorithms for urban traffic forecasting is effective both in terms of observed forecasting accuracy and development, and training efforts. Finally, we discuss problems and obstacles of transferring methodologies and present potential directions for further research. Full article