Enhanced unsupervised domain adaptation through joint constrained classifier and feature learning

Identifying Pathogenesis of Acute Coronary Syndromes using Sequence Contrastive Learning in Coronary Angiography.

Human action understanding serves as a foundational pillar in the field of intelligent motion perception.Skeletons serve as a modality- and device-agnostic representation for human modeling, and skeleton-based action understanding has potential applications in humanoid robot control and interaction. However, existing works often lack the scalability and generalization required to handle diverse action understanding tasks. There is no skeleton foundation model that can be adapted to a wide range of action understanding tasks. This paper presents a Unified Skeleton-based Dense Representation Learning (USDRL) framework, which serves as a foundational model for skeleton-based human action understanding. USDRL consists of a Transformer-based Dense Spatio-Temporal Encoder (DSTE), Multi-Grained Feature Decorrelation (MG-FD), and Multi-Perspective Consistency Training (MPCT). The DSTE module adopts two parallel streams to learn temporal dynamic and spatial structure features. The MG-FD module collaboratively performs feature decorrelation across temporal, spatial, and instance domains to reduce dimensional redundancy and enhance information extraction. The MPCT module employs both multi-view and multi-modal self-supervised consistency training. The former enhances the learning of high-level semantics and mitigates the impact of low-level discrepancies, while the latter effectively facilitates the learning of informative multimodal features. We perform extensive experiments on 25 benchmarks across across 9 skeleton-based action understanding tasks, covering coarse prediction, dense prediction, and transferred prediction. Our approach significantly outperforms the current state-of-the-art methods. We hope that this work would broaden the scope of research in skeleton-based action understanding and encourage more attention to dense prediction tasks.

ARDMR: Adaptive recursive inference and representation disentanglement for multimodal large deformation registration.

Refractive error detection in smartphone images via convolutional neural network.

A novel domain generalization framework for fault diagnosis of rotating machinery based on causal representation learning and causal feature identification

Gd-EOB-DTPA-enhanced MRI radiomics and deep learning models for predicting the pathological differentiation degree in hepatocellular carcinoma.

An attention aided wavelet convolutional neural network for lung nodule characterization.

Multimodal feature learning for gait recognition of camouflaged individuals

Although virtual staining has emerged as a promising alternative to chemical staining through strong feature extraction and color representation capabilities of artificial intelligence, most methods suffer from poor robustness and limited compatibility with the existing clinic workflow. In this Letter, we propose a deep learned label-free virtual staining method to realize accurate and plug-and-play pathological examination with extended depth-of-field by encoding inherent spectral priors into stained visual representations. A custom imaging system with highly flexible optical parameters is constructed for multidimensional pathological spectral information acquisition. An end-to-end supervised spectral stained network (SSNet) is established for efficient spectral cue extraction and accurate stained feature learning. Experimental results across various tissues indicate that the proposed approach achieves robust virtual staining with high morphological accuracy and color fidelity. The proposed method completely gets rid of the utilization of exogenous dyes before imaging, which provides a new panel for fast diagnosis and in-vivo examination.

Deep learning applied to electromyography (EMG) signals enables accurate hand gesture recognition, revolutionizing diverse applications such as human-machine interaction, neural interfaces, and rehabilitative robotics. A well-designed deep learning architecture is crucial for accurately and robustly modeling and decoding the multidimensional information embedded in the EMG data. This survey presents a comprehensive review of state-of-the-art deep learning models and, for the first time, offers a categorization of advanced architectures from the perspective of data representations. EMG, as a distinctive biosignal modality, can be characterized through multiple representational forms, including temporal waveforms, spatial images, spectral domains, and graph-based structures comprising interconnected nodes. Consequently, the optimal model architecture is closely tied to the specific data representation employed. In addition, the limited availability of EMG datasets, particularly those with high-quality labels, remains a critical bottleneck and continues to impede the translation of research advances into widespread real-world applications. We therefore examine emerging semi-supervised and self-supervised learning frameworks, which serve as complementary approaches to fully supervised paradigms. Finally, we outline promising future directions for the development of generalizable and robust deep learning for practical EMG decoding.

A feature-based knowledge distillation (FKD) for offline signature feature learning without signatures

Semi-LLIE: Semi-supervised contrastive learning with Mamba-based low-light enhancement.

Global frequency-aware multi-scale feature learning for point cloud normal estimation

Spectral-constrained global and local feature learning for hyperspectral anomaly detection

A dual-stream regional feature learning and adaptive fusion method for electroencephalogram-based emotion recognition

Enhanced genetic algorithm-optimized deep learning features for lung cancer classification

Rare event classification in multivariate time series is a critical yet challenging task across different industries. Traditional methods often struggle to capture the nonlinear and nonstationary dynamics inherent in complex multivariate time series data, limiting their ability to accurately detect rare events. To address these limitations, this study introduces the Vector Visibility Graph-based Graph Attention Network (VVG-GAT) framework, a novel approach that leverages the power of Vector Visibility Graphs (VVG) to represent multivariate time series as directed, weighted graphs. By encoding both temporal and inter-variable dependencies, the VVG facilitates the application of advanced graph neural networks, particularly the Graph Attention Network (GAT), to address the challenges of rare event classification and early detection. The proposed framework was evaluated on a real-world case study involving a pulp-and-paper manufacturing process characterized by rare paper break events. Experimental results demonstrated that the VVG-GAT framework significantly outperformed traditional models across several metrics. The study further highlights the potential of incorporating VVG-derived network statistics as additional features for machine learning and deep learning models. The VVG-GAT framework represents a significant advancement in rare event classification for multivariate time series generated from complex systems, providing a new solution with broad applicability across various domains.

This paper presents Vi-mCALNET, a multi-constraint active learning-assisted deep-ensemble spatio-textural feature learning model for violence detection in surveillance videos. Unlike traditional approaches, Vi-mCALNET integrates active learning-driven frame selection with deep ensemble learning to enhance classification accuracy while reducing computational complexity. Traditional deep learning vision models for violent crime detection face limitations including inability to use contextual details, long-term dependency issues, gradient vanishing, and accuracy degradation. Vi-mCALNET addresses these challenges through a comprehensive approach. The model employs GLCM, ResNet101, and DenseNet121 for feature extraction, followed by a heterogeneous ensemble classifier comprising SVM, DT, k-NN, NB, and RF. Extracted features are fused into a composite feature vector, processed through PCA and z-score normalization to prevent local minima, convergence issues, and overfitting.The heterogeneous ensemble classifier uses maximum voting to classify videos as violent or non-violent. Vi-mCALNET achieved superior performance with 99.51% accuracy, 99.32% precision, 99.36% recall, and 0.994 F-measure on publicly available datasets.Ablation studies and statistical significance analysis confirmed Vi-mCALNET's robust performance with lower variance, making it suitable for real-time, scalable surveillance applications while reducing annotation costs and computational demands.

Skin diseases represent a significant global health challenge; however, precise automated detection of cutaneous lesions remains difficult due to high intra-class variability, inter-class similarity, and severe class imbalance across disease categories. This paper presents a multi-class skin lesion classification framework based on transfer learning, which integrates an EfficientNet-B3 backbone with a Convolutional Block Attention Module (CBAM) to enhance the learning of discriminative features. EfficientNet-B3, pre-trained on large-scale natural image datasets, serves as a powerful feature extractor, while CBAM improves feature representation by adaptively emphasizing informative channels and spatial locations. This enables the network to focus on diagnostically relevant lesion regions while suppressing background artifacts. The proposed model is trained and evaluated on the DermNet-23 dataset, comprising 23 clinically significant skin disease classes. To address the challenges of multi-class classification and class imbalance, performance is assessed using standard metrics including accuracy, precision, recall, F1-score, and area under the receiver operating characteristic curve (AUC). Experimental results demonstrate that the EfficientNet-B3 + CBAM model achieves 87.1% accuracy, 85.6% macro-F1 score, and 0.94 AUC, outperforming baseline CNN, ResNet50, MobileNetV3, and standard EfficientNet-B3 models. These results highlight the effectiveness of attention-guided transfer learning for developing robust and scalable computer-aided diagnostic systems for skin lesion classification.