ABSTRACT Fatty liver disease (FLD) is a prevalent chronic condition that can progress to clinically significant liver injury if left untreated. Given its high prevalence, FLD poses a substantial public health burden. Traditional Chinese medicine suggests that the appearance of tongue reflects hepatic and metabolic status, enabling artificial intelligence‐based tongue diagnosis to provide rapid low‐cost prescreening for FLD severity. In practice, the development of high‐quality models is hindered by costly disease labelling, scarcity of positive cases, and severe class imbalance. In this work, we propose DualSS, a dual‐diffusion synthetic sampling framework to mitigate these limitations. DualSS first trains a disease‐agnostic dual‐diffusion generator (DDG) on tongue‐image data, leveraging abundant negative samples to learn unified tongue representations. Building on DDG, we design a latent space synthetic sampling pipeline that synthesises realistic positive‐class images to construct a more balanced augmented training dataset. We then couple this pipeline with a noise‐robust training strategy on the augmented dataset to further enhance the learning of downstream diagnostic models. Extensive experiments on FLD severity classification using tongue images and physiological indicators show that DualSS consistently outperforms existing rebalancing baselines, offering an effective and generalisable solution to alleviate positive sample scarcity in tongue image‐based medical diagnosis.
ABSTRACT Handwritten Dongba Character Recognition (HDCR) contains a large number of visually similar characters with subtle and fragile edge cues, posing severe challenges to feature learning. To address this issue, an Edge Channel Aggregation Network (EdgeCANet) model is proposed. The core of EdgeCANet is the Edge Channel Attention Aggregation (ECAA) block, which integrates the Edge‐Guided Enhancement Module (EGEM) for edge‐aware spatial refinement and the Channel Texture Recalibration Module (CTRM) for channel‐wise texture recalibration. Besides, the two‐stage misclassified sample weight redistribution fine‐tuning scheme based on Exponential Moving Average (EMA) algorithm is proposed to stabilise parameter updates, and to emphasise difficult and highly similar classes. The recognition accuracy of EdgeCANet significantly outperforms state‐of‐the‐art models across the constructed high‐similarity datasets HS‐DB and HS‐S20 K, as well as the public DB1404, OBC306, Sketch‐20 K, and ImageNet‐Sketch datasets. The results validate that EdgeCANet effectively enhances recognition accuracy for highly similar handwritten characters and ancient minority‐script characters, offering more promising applicability for related recognition tasks.
ABSTRACT Efficient autonomous navigation in complex and unstructured environments without pre‐existing maps remains a significant challenge for mobile robotics. Drawing inspiration from rodent neural architectures, this study proposes a brain‐mimetic mapless navigation framework for mobile robots. It integrates three key components to achieve autonomous navigation: first, a visual stream‐based localisation model utilising object‐vector cells to compensate for cumulative path integration errors; second, a navigation‐guiding model where the hippocampal‐prefrontal circuitry outputs initial paths through exploration and learning, which are then dynamically optimised for obstacle avoidance through a self‐organising strategy of hippocampal CA1 place cells incorporating boundary‐vector cells' inputs; and finally, the consolidation of these optimised paths into stable navigation habits via spatial cells' theoretical firing rate and spike‐timing‐dependent plasticity learning rule. Extensive 2D and 3D simulation experiments demonstrate that the proposed framework not only outperforms various baseline algorithms in terms of convergence speed and path efficiency but also exhibits robust error correction under motion noise, rapid habit reshaping during task changes and strong invariance to the initial heading direction. Moreover, real‐world experiments on a mobile robot in complex indoor environments further confirm its practical feasibility and effectiveness, which offers a biologically plausible and computationally efficient navigation solution.
ABSTRACT Knee joint moment estimation is a critical component in biomechanical analysis, with profound implications for rehabilitation assessment and the development of assistive devices like exoskeletons. However, existing data‐driven approaches often rely on highly complex network architectures in pursuit of high accuracy, resulting in inherently inefficient training procedures. Furthermore, incorporating new subject‐specific data typically requires retraining the entire model, which substantially increases computational cost. These limitations make it challenging for current data‐driven methods to achieve personalised knee joint moment estimation. To address this challenge, this paper proposes a novel framework based on an Incremental Broad Learning System (IBLS) that can achieve efficient model training through a flattened network architecture and an analytical ridge regression solver. More importantly, it integrates an incremental learning algorithm that enables rapid model updates using only newly acquired subject‐specific data, eliminating the need for costly full retraining. These capabilities collectively make efficient personalised knee joint moment estimation feasible. Extensive experiments on Dataset A and Dataset B demonstrate that our method achieves higher prediction accuracy compared to standard baselines (ANN and LSTM). Furthermore, our approach exhibits competitive performance when compared with advanced deep learning architectures (Transformer and TCN‐LSTM). Notably, the training efficiency is significantly enhanced: the training time required by our method is approximately 30% of that for ANN, 7.5% for LSTM, and only 3.5% for both the Transformer and TCN‐LSTM. Further experimental evidence indicates that incorporating incremental learning provides an additional 60% reduction in training time relative to full retraining, without sacrificing prediction accuracy. In addition, we systematically identify optimal Inertial Measurement Unit (IMU) configurations that balance accuracy with practical wearability, providing actionable guidelines for implementation. This work provides an efficient, accurate, and personalised solution for knee joint moment estimation, paving the way for the development of adaptive and efficient wearable robotic systems.
ABSTRACT Machine learning (ML) integration is becoming increasingly popular in advancing the 4th industrial revolution, known as Industry 4.0. This review paper examines ML applications in CO2 management stages: emission, capture, and conversion. ML models, including multiple linear regression (MLR), multiple nonlinear regression (MNLR), and artificial neural networks (ANN), are used in the emission phase to predict the CO2 emissions for the countries and sectors, variables such as energy consumption, fuel usage, gross domestic product (GDP), population, and other miscellaneous socio‐economic variables. In the capture stage, ML forecasts post‐combustion and direct air CO2 capture efficiencies using models such as ANN, generalized regression neural networks (GRNN), backpropagation neural networks (BPNN), and group method of data handling (GMDH), considering input variables such as catalysts, process parameters, and rheological properties. In the conversion stage, ML models, primarily ANN, multi‐layer perceptron (MLP), particle swarm optimization (PSO), and genetic algorithm (GA), forecast and optimize CO2 transformation into products such as methanol, C2 hydrocarbons, syngas, methane, lower olefins, and biofuels, considering factors such as catalyst compositions and process conditions. This study identifies several trends, with ANN emerging as the predominant ML technique. The discoveries will benefit academia and industry in the development of CO2 management models.
ABSTRACT Flood detection is essential for real‐time applications, including disaster management, emergency response, and alerting people in flood zones. For successful flood detection, accurate flood region segmentation is essential. However, the flood region segmentation is challenging due to the complex background and occlusions with debris and the effect of external adverse factors. In contrast to existing models that focus on satellite and remote sensing images, for real‐time applications, this study proposes a new Transformer‐based Multi‐Scale‐Attention U‐Net (TMSA‐Net) model for flood region segmentation in images with cluttered backgrounds. The proposed model extracts features hierarchically using a Convolutional Neural Network (CNN) encoder. The deepest features are fed to an adapted vision transformer for extracting global context. In the decoder, Multi‐Scale Attention Gates are used in skip connections to selectively filter encoder features before fusion. The proposed model is compared with state‐of‐the‐art methods, including Swin‐UNet, DeepLabV3+, TransUNet, UNet++, and Attention U‐Net. Experimental results show that TMSA‐Net achieves a test IoU of 92.01%, with improvements of 2.82% and 5.95% IoU over DeepLabV3+ and TransUNet, respectively, demonstrating the effectiveness of the proposed approach. The newly created experimental dataset, with ground truth at the pixel level, will be publicly available (released upon acceptance).
ABSTRACT Adverse haze conditions introduce complex degradations that obscure scene details and distort structural cues critical for object detection, posing persistent challenges for vision‐based sensing systems. Although existing haze removal methods have achieved notable improvements in visual clarity, their optimisation objectives are often misaligned with downstream detection requirements, leading to limited detection performance in real‐world scenarios. To address this issue, this work proposes a task‐aligned weakly supervised haze removal framework, termed Dehaze4Detection, which explicitly aligns low‐level restoration with high‐level detection objectives. The framework incorporates a Semantic‐Aware Multi‐Scale Fusion Module (SMFM) that embeds pixel‐level semantic knowledge into the dehazing process, enabling selective enhancement of object regions while suppressing background over‐enhancement to achieve detection‐friendly restoration. To further improve feature quality, a Contrast Self‐Correction Attention Module (CSAM) is introduced, which exploits contrast maps derived from pseudo‐dehazed images to guide attention towards object contours and salient structures. The proposed SMFM and CSAM are integrated into a unified pipeline and jointly optimised using detection and dehazing losses. Extensive experiments on both synthetic and real‐world benchmarks demonstrate that Dehaze4Detection achieves competitive perceptual quality while significantly improving detection accuracy.
ABSTRACT Accurate medical image segmentation with low model complexity remains difficult because lesions are often small in scale and boundary cues are easily corrupted by noise. Although recent segmentation methods have achieved strong performance, many of them rely on increasingly complex architectures with high computational costs, limiting their applicability in resource‐constrained settings. To alleviate this issue, we propose PLNet, a lightweight hybrid segmentation framework for medical image analysis. The network improves representation learning by integrating fine‐grained local‐structure perception with long‐range contextual modelling, thereby enhancing the representation of complex medical images. In addition, a feature selection mechanism is employed to emphasise informative feature responses and suppress redundant activations while preserving discriminative representations. Extensive experiments on multiple datasets demonstrate that PLNet achieves competitive IoU and Dice scores with relatively low model complexity. Overall, these results highlight the potential of PLNet as an effective lightweight framework that balances segmentation performance and computational complexity.
Achieving high accuracy and synergy remains extremely difficult for multi-robot embodied visual localisation, which suffers from persistent real-world challenges such as viewpoint ambiguity, appearance variation and dynamic occlusion. Conventional optimisation-based methods often lead to incorrect feature matching without domain adaptation, whereas existing deep learning models are limited in both precision and generalisability for pose regression. Although semantic information can offer a cognitive-level scene representation, its potential has not been fully exploited due to inconsistent labelling and scaling. To address these issues, this paper proposes a novel semantic and synergic consensus framework (termed SSC-MR) that enhances image retrieval and pose estimation by leveraging multi-granular semantics and multi-robot consensus. Specifically, the spatial topology is attentively generated as a global scene graph for fast image retrieval across heterogeneous robots in a holistic workspace, whereas a local semantic-guided pose estimation strategy refines matched keypoints under semantic consistency constraints. To further improve the efficiency of multi-robot collaboration, an intra-robot minimum redundancy consensus module selects informative keyframes and compresses feature representations to reduce perceptual redundancy, and an inter-robot maximum correlation consensus module further establishes high-confidence state associations among interacting robots for knowledge sharing. Extensive experiments on five public datasets demonstrate that the proposed SSC-MR obtains superior effectiveness and robustness across diverse indoor and outdoor scenes, providing a reliable and scalable solution for embodied visual localisation in practical robotic applications. Notably, it achieves a 62.1% improvement in localisation accuracy with nearly 22 & times; faster speed compared to a typical distributed architecture for 10 collaborative robots. Code is available at: https://github.com/wwsbot/SSC-MR.
ABSTRACT Fine‐grained alignment is crucial for text‐based person retrieval, which searches for relevant pedestrian images using a text query. However, background clutter and semantically vacuous words can cause interference and misalignment, hindering retrieval performance. Existing methods employ masked language modelling to randomly mask and predict words, or leverage multimodal large language models to generate diverse and pure samples. However, these methods rely on external tools and are not sufficiently stable or controllable. In this paper, we propose a pedestrian‐relevant hierarchical parsing (PHP) module to extract well‐aligned fine‐grained visual and textual features for alignment. First, we design a coarse relevant feature mapping (CRFM) module, which uses learnable unified tokens to project both modalities into a shared low‐dimensional space, enabling coarse‐level semantic filtering. Next, we design an expert‐driven feature parsing (EFP) module, which integrates the representational power of a mixture of experts with a modality‐aware gating mechanism to uncover deep semantic associations between text and image features. Extensive experiments on three public large‐scale datasets demonstrate that our approach outperforms existing state‐of‐the‐art methods, achieving, in particular, a 5.68% improvement in mAP and a 4.81% boost in R1 accuracy on the CUHK‐PEDES dataset.
Graph transformers (GTs) with elaborate positional/structural encodings (PEs/SEs) have excelled in graph representation learning, especially in graph-level tasks. However, their potential in large-scale node classification remains untapped for several reasons: (i) Current PEs/SEs are insufficient in modelling large-scale real-world graphs, where a multi-angle portrayal of node properties is required. (ii) The common integration of PEs/SEs with self-attention overlooks nodes' distinct preferences, leading to sub-optimal performance. (iii) The global receptive field of self-attention leads to quadratic complexity with respect to graph size and introduces potential noise, which significantly hinders the learning of node-level tasks. In this paper, we propose the adaptive node property graph transformer (ANPGT) to address the above issues. ANPGT, with its node property extractor (NPE) and adaptive property integrator (API), flexibly extracts multi-angle node Properties and integrates them according to node-specific preferences. Theoretically, we analyse the necessity of adaptive node property learning. Empirically, ANPGT achieves or matches the best performance on nine real-world datasets (up to three million nodes) and various synthetic benchmarks regarding heterophily and noises. ANPGT exhibits almost linear complexity relative to graph size.
ABSTRACT Exploring clean energy alternatives to fossil fuels has become a major research focus worldwide. Wind energy is regarded as one of the most promising renewable energy sources due to its clean and sustainable nature. However, wake interactions among turbines significantly reduce power conversion efficiency, making wind farm layout optimization (WFLO) crucial for maximising energy output. As the number of turbines increases, wake interactions become more pronounced, further deteriorating overall efficiency. Metaheuristic algorithms have been widely adopted to address the complex constraints and design objectives of WFLO. Nevertheless, traditional heuristic methods often suffer from poor solution quality and premature convergence when dealing with large‐scale WFLO problems under complex wind conditions. To overcome these limitations, this study proposes a multi‐strategy synergy‐based differential evolution algorithm (LSDE) for large‐scale WFLO under complex wind scenarios. The proposed method is evaluated against nine representative WFLO algorithms under four complex wind scenarios (4, 5, 6 and 7 wind directions) and three turbine scales (30, 50 and 100 turbines). Experimental results demonstrate that LSDE achieves superior performance, stability and robustness. Specifically, LSDE improves power conversion efficiency by 3.40%, 3.51%, 3.10% and 3.36% under the four wind scenarios, respectively, compared with other mainstream algorithms.
ABSTRACT Load‐frequency control (LFC) comprises a primary process in interconnected electrical power systems, playing a critical role in maintaining the stability and reliability of the electrical grid. It is of paramount importance that the designed controller functions in an optimal manner, particularly with regard to the compression of area frequency and tie‐line power deviations within the scheduled limits. In light of these considerations, this study puts forth the inaugural derivative‐free and delay‐based application of a predictive proportional‐integral (PPI) and its fractional‐order approach (FOPPI) controllers to address the LFC challenge in a PV‐integrated two‐area power system. To attain the most effective controller parameters of the PPI and FOPPI, the artificial rabbits optimisation (ARO) methodology is utilised by aiming to minimise the integral of the time‐weighted absolute error (ITAE) performance measurement. The effectiveness of the proposed FOPPI controller is evaluated through a comparative analysis of existing controller structures in the literature and the classical PPI controller based on reported ITAE values in the literature. The proposed approach is validated through a series of detailed and systematic performance evaluations, such as time‐domain analysis, uncertainty in system parameters, stochastic loading conditions, stability analysis and cyberattack performance of the proposed controllers. Furthermore, a cyberattack detection index for the controllers is considered together with a statistical analysis, and the dataset and simulation models supporting the study are made available as described in the Data Availability Statement. As evidenced by the outcomes of the performance assessments, the FOPPI controller, as conceptualised in this investigation, displays an augmented capacity for dynamic response and resilience in the face of specified scenarios when contrasted with alternative methodologies, including the PPI controller.
ABSTRACT Accurate surface‐pressure prediction over broad operating envelopes is critical for supersonic aerodynamic analysis and design. To overcome the bottlenecks of traditional computational fluid dynamics (CFD) in real‐time performance and computational efficiency, data‐driven deep learning methods have emerged. However, existing data‐driven models face two fundamental physical challenges: first, the pressure response to Mach number, angle of attack and altitude is strongly spatially heterogeneous, which cannot be adequately captured by traditional global condition fusion methods; second, the inherent strong anisotropy of supersonic flows renders standard isotropic graph construction based on Euclidean distance prone to spurious cross‐shock connections, resulting in nonphysical smoothing of predictions. To address these issues, we propose aircraft pressure transformer network (APTNet), a physics‐informed deep learning framework that explicitly encodes flow–geometry interactions and streamwise physical priors. Specifically, we first propose a condition‐sensitive local modulation (CSLM) module, which utilises a surface‐normal–freestream alignment term to distinguish windward and leeward regions, and learns per‐point sensitivity fields to implement physically interpretable local pressure modulation via a feature‐wise linear modulation (FiLM) mechanism. Furthermore, we introduce a flow‐aligned anisotropic graph construction strategy, which reshapes the local receptive fields of graph nodes into ellipsoids elongated along streamlines, enforcing information to propagate preferentially along streamlines and suppressing cross‐shock feature mixing. Experiments on diverse complex aircraft configurations demonstrate that APTNet consistently outperforms existing baseline models in terms of both accuracy and efficiency. Beyond quantitative improvements, the model provides interpretable intermediate signals, such as attention maps and feature responses, which can be directly integrated into the aerodynamic design workflow.
Extracting spatio-temporal cues from neighbouring frames is challenging in video super-resolution (VSR). Although deformable alignment-based VSR methods have shown promise in aligning neighbouring frames with the reference frame, most existing methods rely on one or a few traditional convolutions to estimate motion offsets for spatio-temporal alignment, restricting receptive field size and alignment accuracy. To address these limitations, we propose an effective spatio-temporal alignment network (ESTA-Net) for VSR. The core component of our method is the group convolution-based alignment module (GCBAM), which utilises cascaded group convolutions to learn offsets across both the original and downsampled resolutions. By employing group convolutions rather than traditional convolutions, GCBAM enables the deformable alignment to achieve a wider receptive field with lower computational cost, thereby improving the accuracy of offset estimation. Additionally, the bi-scale alignment strategy within GCBAM enhances robustness to complex and large-scale motions. Furthermore, we introduce an attention-based feature enhancement module (AFEM) to refine the aligned features, focusing on critical details to improve reconstruction quality. Extensive experiments on standard benchmarks show that our ESTA-Net achieves superior VSR performance against other advanced methods, while maintaining a good equilibrium between model size and performance.
Open-vocabulary 3D querying based on 3D Gaussian splatting (3DGS) shows great promise in facilitating accurate 3D query capabilities of AI systems. These methods typically rely on pre-captured multi-view images to enable natural language interactions with 3D scenes. In practice, when embodied AI encounters unexplored scenes, it is difficult to obtain observations from different viewpoints beforehand. This challenge highlights the importance of exploring natural language-driven 3D scene querying from a single current viewpoint. This paper proposes single view language Gaussian splatting (SVLGaussian) for the novel task: Open-vocabulary 3D querying based on the input single view. By leveraging multi-round inference of multimodal large language models, SVLGaussian efficiently generates pixel-level semantic probabilities and rapidly embeds them into a 3D Gaussian field, enabling real-time language-guided semantic querying. To verify our model, we annotated three datasets: Lerf_ovs and 3D-OVS, which are tailored for open-vocabulary 3D querying, and RE10K, which is adapted for single-view 3D reconstruction. Both quantitative and qualitative results show that our method effectively supports open-vocabulary 3D querying from a single view.
ABSTRACT As an attestation engagement, auditing is required to provide reasonable assurance for its conclusions. Traditional auditing has limited capacity to handle unstructured data and is usually based on audit sampling techniques, which can lead to the neglect of important audit evidence during the auditing process and result in a higher audit risk, making the assurance level uncontrollably guaranteed. Meanwhile, the nature of attestation risk, which is characterised by multi‐factor triggering chain transmission and time evolution, brings difficulties to the acquisition of critical audit evidence. Therefore, this paper realizes the Full‐chain Loop Tracking Audit through the business cycle audit process of ‘Obtaining Evidence of Internal Sequential Timing Audit—Obtaining Evidence of Horizontal Expansion of Audit Matters—Penetrating the Multi‐Dimensional Evidence Association Rules—Achieving the Anchor Point at The Reasonable Assurance Level’. This is called the Full‐chain Loop Tracking Auditing framework (FcLTA framework). Firstly, ‘Obtaining Evidence of Internal Sequential Timing Audit’ is a process of identifying and obtaining abnormal trends in accounting book facts. This process obtains ‘Time Series Audit Evidence’ by embedding the time characteristics of audit evidence in the risk feature vector within the table. Secondly, ‘Obtaining Evidence of Horizontal Expansion of Audit Matters' is a ‘Triangular Verification’, system constructed by integrating financial, business and public opinion data. This process provides ‘Horizontal Expansion Audit Evidence’ by identifying isolated risk signals in financial information to address issues such as discrepancies between accounts and reality and performance whitewashing. Thirdly, ‘Penetrating the Multi‐Dimensional Evidence Association Rules' is a process of integrating audit evidence and resolving conflicts. This process generates ‘Multi‐Dimensional Penetration Audit Evidence Chain’ by constructing a high‐order association across modalities and dimensions for verifying risks. Finally, ‘Achieving the Anchor Point at The Reasonable Assurance Level’ is a testing process for detecting risks with the goal of obtaining high‐level assurance of audit evidence. It mainly includes two processes: predicting the level of audit assurance and determining whether to continue the audit. This paper has implemented the full‐chain circular tracking audit through the FcLTA framework, and the generated audit evidence can meet the high‐level assurance requirements of the attestation engagements, reducing the risk of audit failure.
ABSTRACT Traffic Flow Forecasting (TFF) is a foundational task in the development of Intelligent Transport Systems (ITSs). The primary challenge is to undertake a comprehensive exploration of the intrinsic dynamic spatiotemporal correlations of the road network, unveiling the long‐term evolutionary traffic trends. Furthermore, most existing methods often solely depend on the single traffic condition and neglect the enhancement of correlated features collected from traffic sensors in prediction. To this end, we propose a dynamic correlation‐information‐fusion‐based (DCIF) spatiotemporal network for TFF, which models the spatiotemporal correlations of road networks, thereby effectively capturing dynamically changing characteristics. Specifically, a spatiotemporal feature enhancement (STFE) mechanism is employed to capture the directional and location‐aware characteristics of traffic flow, thereby enhancing the representation of traffic flow and the capability of spatiotemporal feature extraction. Then, a gated attention unit (GAU) is constructed to meticulously extract the deep dynamic trends inherent within traffic data. Finally, a dynamic feature matrix (DFM) is formulated, incorporating spatial graph convolution to provide comprehensive semantic contextual information. The DFM captures the dynamic topology of the deeper feature network in real time by fusing spatial node information and traffic speed features as correlation information. Extensive experiments demonstrate that DCIF significantly outperforms other baselines in prediction accuracy, thereby further substantiating its validity and reliability in TFF.
