Facial expressions convey abundant emotional cues and serve as one of the primary means for humans to communicate emotions. However, existing Facial Expression Recognition (FER) methods generally suffer from expression ambiguity (high inter-class similarity and low intra-class diversity), which significantly limits FER performance. Although recent studies have attempted to alleviate these issues by introducing attention mechanisms and Transformer architectures, most approaches still lack explicit awareness of semantically critical regions. In addition, they remain sensitive to imbalanced sample distributions in FER datasets, resulting in poor interpretability and robustness. To address the aforementioned issues, this paper proposes a Vision Transformer with Visual Awareness Positional Encoding. Specifically, VAPE first a visual awareness position encoding is proposed to deeply integrating expression semantics with position encoding, thereby overcoming the limitation of traditional position encoding, which only provides generic location information without distinguishing semantically critical regions. Finally, a semantic embedding classification module is designed to achieve accurate and semantically consistent recognition of different expressions by measuring the alignment between image features and the semantic embeddings of expression labels, ultimately enhancing the model's interpretability and robustness. We perform experiments on three public datasets to validate the effectiveness of the proposed method, achieving accuracies of 91.45%, 66.81%, 63.34%, and 62.30% on RAF-DB, AffectNet (7cls), AffectNet (8cls), respectively.

To address the challenges of insufficient continuous learning and catastrophic forgetting in maritime satellite communication network intrusion detection systems (IDSs) when responding to new attacks, this paper proposes a dynamic extension based continuous learning model (DECL). The DECL model innovatively integrates two core mechanisms: label mapping and dynamic expansion. The label mapping mechanism uses a function to scalably map new attack categories to new output nodes without modifying the existing structure. Furthermore, when learning new intrusion types, the dynamic expansion mechanism creates new nodes and connects them to the existing tree structure, rather than modifying already trained nodes. This ensures the integrity of historical knowledge and effectively mitigates catastrophic forgetting. To validate the model in a real maritime environment, we constructed a complete intrusion detection system that aggregates traffic data, extracts statistical features, and feeds structured data into the DECL model for analysis. Experiments were conducted on real-world ship satellite communication data and the CICIDS2017 dataset. Results show that DECL surpasses four competing algorithms, including EWC and GEM, in terms of flexibility and stability, demonstrating excellent generalization and anti-interference capabilities. In continuous learning tasks, the model’s knowledge retention rate always remains above 90

Detecting ships from maritime remote sensing imagery remains a challenging task due to several factors, including weak representation of small targets, interference from complex backgrounds, and large-scale variations in object size. Existing CNN-based backbone networks often rely on fixed feature generation patterns, resulting in limited representation capability. Moreover, most attention mechanisms lack effective collaborative modeling between spatial and channel dimensions, making it difficult to enhance small targets and suppress background noise. Additionally, the local receptive field of convolution limits the modeling of long-range dependencies. To overcome these limitations, this paper proposes a novel backbone network, termed DASNet. An adaptive derived feature generator is first introduced to dynamically regulate the generation of feature maps, enabling more effective extraction and enhancement of weak features while reducing redundancy. Then, to enhance small ship visibility in complex backgrounds, a dynamic multi-semantic joint attention is incorporated early to achieve spatial-prior-guided channel recalibration. Finally, in the later stages, we embed a Vision Mamba-based global interaction module, which models long-range dependencies under linear computational cost, providing essential contextual support for accurately distinguishing small targets from their surroundings.

Multiple vessel tracking plays a vital role in maritime surveillance systems. Previous studies have typically integrated object detection and trajectory association techniques to address this problem, but they still face some significant challenges. On one hand, these methods are susceptible to losing tracked targets due to long-term occlusion by other obstacles or slow-moving vessels in inland waterways. Moreover, traditional models encounter difficulties in accurately capturing the global appearance features of the vessels in images, which leads to a decline in vessel detection performance. To address the issues above, this paper proposes a novel Vessel Status Augmented Track (VSATrack) framework for multi-vessel detection and tracking. Specifically, we present a Motion-Matching Optimization Module (MMOM), which handles long-term occlusion through identity matching between consecutive frames. Besides, a vessel feature enhancement module (VFEM) with several residual convolutional layers and channel reconstruction units (CRU) is designed to effectively capture the vessels features in complex inland waterway backgrounds without introducing redundant channel information. Finally, a bidirectional feature pyramid network (BiFPN) is utilized to fuse vessel appearance features from different scales, enhancing the capability to learn cross-scale features of vessels to some extent. Experimental results demonstrate that our VSATrack method outperforms the state-of-the-art methods, particularly in reducing the number of vessel ID switches (IDSW).

This comprehensive review examines the works of reinforcement learning (RL) in ship collision avoidance (SCA) from 2014 to the present, analyzing the methods designed for both single-agent and multi-agent collaborative paradigms. While prior research has demonstrated RL's advantages in environmental adaptability, autonomous decision-making, and online optimization over traditional control methods, this study systematically addresses the algorithmic improvements, implementation challenges, and functional roles of RL techniques in SCA, such as Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Multi-Agent Reinforcement Learning (MARL). It also highlights how these technologies address critical challenges in SCA, including dynamic obstacle avoidance, compliance with Convention on the International Regulations for Preventing Collisions at Sea (COLREGs), and coordination in dense traffic scenarios, while underscoring persistent limitations such as idealized assumptions, scalability issues, and robustness in uncertain environments. Contributions include a structured analysis of recent technological evolution, and a Large Language Model (LLM) based hierarchical architecture integrating perception, communication, decision-making, and execution layers for future SCA systems, which prioritizes the development of scalable, adaptive frameworks that ensure robust and compliant autonomous navigation in complex, real-world maritime environments.

In inland waterways, vision-based maritime vessel tracking technology is crucial for maritime safety. Due to varying vessel speeds and angles, tracking vessels in complex environments poses a significant challenge. Currently, the mainstream video tracking method is the “Tracking-by-detection” (TBD) approach, which can meet basic tracking requirements. However, its reliance on shape-based recognition leads to bulky models and challenges in improving accuracy. To address vessel tracking confusion in complex scenarios, we introduced ship motion direction parameters into the Hungarian matching algorithm. Specifically, by combining IOU with ship motion direction, we achieved an efficient tracking mechanism that reduces computational costs. Additionally, by integrating YOLOX with the matching module, we improved tracking efficiency. Our proposed method demonstrates excellent performance on an inland waterway video dataset, showing significant improvements in MOTA compared to other methods, along with a noticeable reduction in ship ID switches.

Space-Based Infrared Tiny Ship Detection (SBITSD) aims to accurately identify tiny ships from images captured by Earth-orbiting satellites, facing challenges such as long imaging distances, small target sizes, and low visual contrast between ocean backgrounds and foreground targets. To address the SBITSD task, this paper proposes a novel deep learning model, Dual Swin-Net, based on an optimized MTU- Net, to improve detection accuracy and reduce the false alarm rate. The model employs a Dual-Stream Lightweight Swin Transformer (DSLST) mechanism to extract local features of tiny ships from multi-level features and capture long-range relationships between targets and backgrounds. Additionally, a Hybrid Multi-level Decoder (HMD) is introduced to perform multi-level feature fusion and generate the final prediction feature map. The model first extracts multi-level features through the Multi-level Feature Extraction (MFE) module, then utilizes DSLST to capture long-range relationships, and finally decodes the results of multi-level features and long-range relationships between targets and background through HMD to achieve prediction feature map. Experiment on the NTUDT-SIRST-Sea dataset demonstrates that Dual Swin-Net outperforms the existing MTU-Net model in terms of detection probability, achieves a lower false alarm rate at high probability thresholds, and exhibits a higher average IoU, showcasing its superior performance.

Visible-infrared cross-modality pedestrian re-identification (Cm-ReID) devotes to mapping the pedestrian photographs of the same identity from different cameras. Despite significant achievements in previous CNN-based works, there remain two limitations: (i) cross-modality feature mismatch caused by data heterogeneity and (ii) intra-modality feature variations resulting from different human postures and shooting angles. To overcome these limitations, in this paper, we proposed a Cm-ReID model based on dual-constraint capsule network (DCCN) to explore feature representations among modalities. Specifically, to cope with cross-modality feature mismatch, we design a modality mitigation module (MMM). It utilizes channel attention mechanism, extracts discriminative features from feature maps to better mitigate the modality discrepancy. With channel attention mechanism, the DCCN can better perceive identity information on both modalities. Furthermore, to address intra-modality feature disparity, we design dual-constraint mechanism in DCCN, which consists of intra-class aggregation module (CAM) and inter-class sparse module (ISM). The former is proposed to aggregate the features of the same pedestrian under different perspectives, and the latter is to sparse the features of different pedestrians. Comprehensive experiments on the public SYSU-MM01 and RegDB datasets reveal the DCCN’s superiority over current practices.

Accurate prediction of vessel trajectories in coastal areas poses a significant challenge due to the large number of irregular trajectories. Existing trajectory prediction studies predominantly employ recurrent neural network (RNN) and Transformer-based methods. However, the former often encounter challenges such as gradient vanishing or exploding, and the latter tend to focus on global temporal dependencies, making it difficult to capture local irregular trajectory features in coastal maritime areas. Recently, graph-based methods have also been used to predict trajectories, however processing graph-structured data introduces significant increase in computation. In responding to these, this paper proposes a framework based on a novel lightweight Slice-Diff self attention, which consists of several key components. Firstly, the trajectory slice difference encoder (TSDE) utilizes slice embedding (SE) to enrich the cross dimensional dependencies contained in the input sequence, and then combines Slice-Diff self attention (SDSA) and fine-grained convolution (FGC) to comprehensively capture sequence-specific positional and directional information. Additionally, an auxiliary model, stepping bidirectional long short-term memory (S-BiLSTM) is developed to capture global temporal dependencies within the whole sequence. Finally, the fine-grained trajectory features obtained from TSDE and the global temporal dependencies compensated by S-BiLSTM are combined through the fully connected layer to predict coastal vessel trajectories. Extensive experimental results on three real-world automatic identification system (AIS) datasets demonstrate the effectiveness of proposed framework against other baselines.

Dense video captioning (DVC) involves detecting each event from untrimmed videos and generating corresponding descriptions for them, thus encompassing both the Event Localization (EL) and Event Captioning (EC) tasks. For event localization, previous methods directly decode the encoded features to generate temporal event proposals (TEPs), resulting in a reduced Utilization Rate (UR) of certain features. We propose a Progressive Feature Refining (PFR) module, which progressively augments underused or reduces over-accessed features to the decoder, allowing the model to identify events within a given time frame and establish connections between them. To our knowledge, this is the first exploration of such an approach in the DVC domain. For event captioning, while recent methods utilizing multi-scale deformable attention (MSDA) have significantly improved description quality, they tend to overlook detailed and comprehensive scene content, focusing solely on salient objects, which compromises the narrative coherence of the generated captions. We present a novel Deformable-Guidance Describer (DGD) module, which ensures that salient information from MSDA interacts with overall context before generating each word. Without bells and whistles, extensive experiments on ActivityNet Captions and YouCook2 datasets demonstrate that Our method achieves state-of-the-art performance in event localization compared to the latest CLIP-based models, and is on par with them in event captioning while requiring only about 4.6% of their parameters.

Learning the intricate spatio-temporal interactions among vessels poses a fundamental challenge to vessel trajectory prediction (VTP). Prevailing graph neural network-based methods typically employ a three-stage independent architecture, modeling spatial and temporal dependencies separately, which not only overlooks the temporal continuity of spatial interactions but also introduces higher computational overhead. Moreover, considering interaction semantics at distinct moments equally during the modeling process may restrict the effective learning for certain key features. To address these issues, our paper proposes an innovative Heterogeneous Spatio-temporal Graph Neural Network (HSTGN) that employs hierarchical aggregations, encompassing vessel node-level, temporal edge-level, and self-attention-level aggregations, tailored for VTP. Specifically, a heterogeneous graph is first constructed through the transformation of node and edge generators, allowing vessels to interact freely with each other across temporal dimensions. A key innovation is the vessel multi-relational graph aggregation module (VMGA), which selectively aggregates edge-specific semantics and assigns adaptive weights to enhance interaction modeling. Furthermore, we devise a gated temporal convolution module (GTCN) to compensate for potential global dependencies that VMGA might not capture. Finally, HSTGN integrates VMGA and GTCN in parallel to refine spatio-temporal correlations for more precise prediction. This study enhances VTP accuracy and efficiency, enabling safer navigation and optimized maritime route planning. Extensive experiments on three real-world Automatic Identification System (AIS) datasets demonstrate that HSTGN outperforms state-of-the-art approaches, achieving a 9.47% improvement in mean absolute error (MAE), 1.43% in root-mean-square error (RMSE), 2.72% in mean absolute percentage error (MAPE), and a 43.06% increase in inference speed, while utilizing fewer model parameters.

This scoping review examines the integration of digital and intelligent technologies in early childhood education (ECE) from 2010 to 2024, analyzing 29 selected empirical studies, focusing on algorithmic approaches and logic for ECE. While prior research has emphasized data-driven outcomes, pedagogical practices, and ethical challenges, this study addresses a critical gap by systematically analyzing the underlying algorithms, implementation forms, and functional roles of intelligent technologies such as Internet of Things (IoT), machine learning (ML), virtual reality (VR), data mining, and robotics. The review highlights how these technologies address challenges in ECE, such as health diagnostics, personalized learning, safety, and equity while underscoring persistent limitations, including security vulnerabilities, insufficient interdisciplinary integration, and a lack of longitudinal efficacy studies. Contributions include a taxonomy of algorithmic functions and implementation patterns, offering stakeholders foundational insights for evidence-based technology adoption. Future directions may include analyzing recent technology advancement in ECE, for example, “digital twins” for child development, AI-driven adaptive systems, and privacy-preserving frameworks using federated learning. The study advocates for strengthened security protocols, interdisciplinary curricula, and mixed-methods evaluations to advance ethical, child-centric innovations. By bridging technological and pedagogical perspectives, this work informs policy and practice for leveraging digital tools to optimize early learning outcomes while mitigating risks.

Lightweight anchor-free detection frameworks have gained prominence in resource-constrained maritime surveillance scenarios. Despite significant advancements in previous research, two major limitations persist: (i) existing methods often lead to matching bias when assigning fixed threshold samples, which has a negative impact on the convergence and generalization of the model; (ii) the lack of coordination between classification and positioning tasks restricts the improvement of detection accuracy. To address these challenges, this paper proposes a lightweight anchor-free detection framework (LAFNet) based on dynamic allocation guidance. LAFNet enhances detection performance through optimized feature extraction and sample allocation strategies while improving task synergy via a refined loss function. Experimental results demonstrate that LAFNet achieves an optimal balance between accuracy and efficiency while maintaining lightweight characteristics, offering an efficient and robust solution for real-time SAR ship detection.

Rapid ship detection in large sea areas is crucial for maritime transportation safety. While deep learning methods excel in high-resolution (HR) remote sensing ship detection, medium-resolution (MR) imagery is better suited for real-time monitoring of waterways and abnormal ship screening due to its broader coverage and higher timeliness. Advancing MR ship detection can mitigate the inefficiencies and high costs of HR monitoring, supporting the development of an intelligent, cost-effective maritime security system. However, research on MR ship detection remains limited, and existing small-ship detection methods have notable shortcomings.(i) Most approaches rely on the Vision Transformer (ViT) for feature extraction, but its limited local information processing and single-feature representation hinder effective multi-scale target detection.(ii) Feature interaction in existing methods is unidirectional: CNNs extract local features, which are then fed into ViT without feedback, preventing full optimization of global semantics and local details.To address these issues, we propose a feature-enhanced ViT backbone with convolutional multi-scale interaction. Additionally, we introduce a CNN-Transformer bidirectional fusion module to enable cross-level multi-scale feature fusion. Experimental results demonstrate that our method outperforms state-of-the-art tiny ship detection techniques.

Versatile Video Coding (VVC) uses Quadtree nested multi-type tree (QTMT) for CU partitioning. Compared with the previous generation coding standard High Efficiency Video Coding (HEVC), it can save about 50% of the bit rate while ensuring the same video quality, but the coding time is greatly increased. In order to solve the problem of high time complexity of Coding Unit (CU) partitioning in VVC, this paper proposes a CU partitioning method based on image texture complexity. This method analyzes the texture of CU and calculates relevant statistics to judge the texture complexity, and then determines the CU partitioning strategy. Experimental results show that the proposed method can reduce the video coding time with little loss of video quality, and has good coding efficiency. Compared with the reference software VTM22.2, the coding time is reduced by 15.44% and the BDBR is increased by 1.94%.

Vessel re-identification (re-ID) is a crucial task in maritime supervision, enhancing maritime safety and improving the maritime situational awareness system. However, distinct from land-based scenarios involving vehicles or pedestrians, vessels, as enormous rigid bodies situated in the dynamic marine environment, face unique challenges such as significant variations in the scale of discriminative features and unpredictable sway. Furthermore, there is a limited number of publicly available datasets for vessel re-ID in complex backgrounds. In this paper, to overcome these challenges, a novel Hierarchical Perceptual Aggregation Network with Inclination-Aware Attention (HPAN-IAA) is proposed. HPAN-IAA comprises two main modules: the Hierarchical Perceptual Aggregation Block (HPAB) and the Inclination-Aware Attention Block (IAAB). Specifically, in HPAB, a hierarchical perceptual function is introduced to decompose visual information of vessels into discriminative features at multiple levels. These feature maps with different levels of detail from diverse network layers are then fused together by concatenation, resulting in a comprehensive feature representation that effectively integrates information across various scales. Conversely, to address the irregular variations and random omissions in discriminative feature distribution caused by unpredictable vessel sway, in IAAB, the Channel Collaborative Attention Module and the Pyramidal Spatial Attention Module are designed to adaptively extract potential discriminative features within each channel and spatial dimension, enhancing model's ability in effectively extracting and utilizing irregularly changing discriminative features. Moreover, we propose a novel vessel re-ID dataset-VesselReID-2258. Extensive experiments conducted on VesselReID-2258 and the publicly available dataset VesselReID demonstrate that HPAN-IAA outperforms the current state-of-the-art methods,achieving superior performance with mean Average Precision scores of 0.861 and 0.823.

Modeling the functional impact of sequence variation is a critical issue for both understanding and developing proteins. An Evolutionary Sequence and Gaussian Mixture Model (ESGMM) for predicting variant pathogenicity is presented in this paper. The model is trained on 2715 clinical proteins and their homologous sequences, using a Transformer-based protein language model to discover evolutionary patterns of amino acids from multiple sequence alignment (MSA). To fully mine deep information of MSA two-dimensional data, an axial attention mechanism is introduced during training. The model estimates the probability of all variants compared to the wild type and calculates variant scores. To categorize variations as pathogenic or benign, a global-local Gaussian mixture model is then constructed for each variant, and ESGMM scores are produced for each variant employing a combination of global and local information. Particle swarm optimization (PSO) is introduced to optimize the local Gaussian mixture model and further quantify the uncertainty of the classification, which enhances the model prediction precision. Experimental results demonstrate the superiority of the optimized ESGMM algorithm in predicting the pathogenicity of variants.

Currently existing vessel trajectory predictions are predicted for individual vessels. However, due to the complexity of the maritime environment, it’s not feasible to capture the mutual information among vessels. In order to better capture the dynamic interrelationships between vessels at sea, inspired by social pooling and message passing mechanism, we design a Social-Passing mechanism to dynamically update the node information of vessels, and combined with LSTM network, we propose a SP-LSTM model. In our model, we first utilize the LSTM network to extract features from the historical trajectory data of a single vessel. Subsequently, we perform the Social-Passing analysis of the vessels at the current time node, based on the distance range between vessels, to update the information of all current nodes. Meanwhile, since multi-vessel datasets are difficult to collect, we propose an AIS Multi-Vessel Data processing algorithm (AMVD) to process the datasets. Finally, we compare several existing models, and all of our SP-LSTM achieves better performance results.

In the age of digitalization and automation, research on data-driven marine engine fault diagnosis is ongoing. However, the volume of normal data in real-world environments significantly exceeds that of fault data, and there will be unknown fault modes that are not collected, which leads to incorrect diagnosis. To address the challenge of open-set fault diagnosis for marine engines under conditions of data imbalance, this paper proposes a hierarchical fault diagnosis method, based on Isolation Forest (iForest), Extreme Learning Machine (ELM) and Parameter-Adaptive Density-Based Spatial Clustering of Applications with Noise (PADBSCAN), named HFD-IEP. To validate the effectiveness of the proposed method, AVL BOOST was utilized to simulate a two-stroke marine diesel engine. The simulation results confirm the method's effectiveness.

In recent years, deep learning techniques especially convolutional neural networks (CNNs), have shown significant progress in facial expression recognition. However, most prior studies are susceptible to the “fuzzy phenomena” in FER datasets, which is a common phenomenon in real-world scenarios. In addition, traditional CNNs are usually sensitive to rotations and pose variations in images, which is a common challenge in facial expression recognition because expressions can occur at different head postures. To address the above-mentioned problems, we propose a Spatial-channel Capsule Aggregated Network with Dynamic data Cleansing Module (SCAN-DCM). Specifically, we incorporate spatial and channel attention mechanism into vanilla capsule network for better capturing the relative position relationships and orientation features among the facial Action Units (AUs), enabling the model to adaptively perceive the facial AUs with high feature information density. Furthermore, the Dynamic data Cleansing Module (DCM) is proposed to address the “fuzzy phenomena” via enhancing the feature expressions of samples with high Contribution Coefficients (CC), suppressing those with low CC, and curing the wrongly labeled samples in the low CC. Extensive experiments are conducted on the FER datasets RAF-DB as well as FERPlus in natural scenes.SCAN-DCM achieves an accuracy of 82.83% on FERPlus and 86.28% on RAF-DB, which shows that our approach has remarkable performance.