Advanced Intelligent Computing Technology and Applications

Visual question answering is a complex task focused on answering questions about images. Current methods exhibit several limitations, including suboptimal multimodal feature matching and ineffective multimodal fusion. To alleviate the above problems, we propose Enhanced Multimodal Alignment and Fusion Networks (EMAFN) to improve the alignment and fusion of multimodal features. We design the Location Attention Module (LAM) to enhance multimodal alignment. This module leverages the location information of objects in images to guide attentional operations, facilitating semantic interaction both within and across modalities. Additionally, we employ a contrastive loss function as a similarity metric to further refine multimodal alignment. To improve multimodal fusion, we design the Attention-based Multimodal Fusion Module (AMFM). This module utilizes aligned modal features obtained from previous steps and employs an attention mechanism to capture highly correlated features between text and images, thereby achieving effective multimodal fusion. A large number of experiments conducted on the VQA-v2 and GQA datasets show that our model achieves better results than current approaches.

Industrial Control System (ICS) network traffic classification is crucial for network security and management but remains challenging due to the imbalanced nature of traffic data, privacy constraints, and the lack of labeled samples. Existing solutions primarily rely on supervised learning methods that require extensive labeled data, limiting their generalization in real-world industrial environments. To address this, we propose PHierT, a Privacy-Preserving Hierarchical Transformer Model that learns hierarchical representations from unlabeled ICS network traffic while ensuring privacy protection by excluding payload data. PHierT pre-trains a deep contextualized traffic representation model on large-scale unlabeled data and fine-tunes it with a small number of labeled samples. By leveraging hierarchical traffic encoding and a novel Masked Packet Prediction (MPP) task, our model achieves stateof-the-art performance across multiple ICS network traffic classification tasks. Experimental results on four real-world datasets demonstrate that PHierT generally outperforms existing methods, achieving a 29.17% improvement in F1-score for intrusion detection on CIC_MODBUS and a 3.25% increase on 2017QUT_DNP3. Notably, our approach effectively preserves industrial privacy while enhancing classification performance, providing new insights into encrypted ICS traffic analysis and generalizable representation learning.

Temporal Graph Neural Networks (TGNNs) are powerful tools for capturing intrinsic interactions among entities in real-world scenarios. However, numerous studies have shown that TGNNs are vulnerable to adversarial perturbations. While robustness certification has been extensively applied to fixed-dimensional models, such as Graph Neural Networks (GNNs), TGNNs have received significantly less attention in terms of robustness analysis. To bridge this gap, we propose an input-dependent randomized smoothing method called IDRS. Specifically, the IDRS method serves as a robustness certification framework tailored for temporal link prediction (TLP) models against modification and injection attacks. Furthermore, we theoretically certify the robustness radius of a TLP model when subjected to perturbations measured by edit distance. Experimental results on the WIKI, REDDIT, and MOOC datasets demonstrate that the proposed method achieves higher clean accuracy than the baseline while maintaining comparable robustness certification performance.

The main challenge of multimodal sentiment analysis (MSA) is to effectively integrate and optimize information from diverse modalities, such as textual, visual, and acoustic. This integration is crucial for achieving more accurate analysis and comprehension of human emotional states. However, the majority of previous studies have not fully extracted and utilized the salient information in the input data, nor have they thoroughly explored the intricate connections between different modalities, thus failing to accurately identify human sentiment. To tackle this challenge, we propose a framework named Mutual Information and Cross-modal Text Enhancement (MICTE) to resolve the modal fusion problem. First, we propose a cross-modal text-enhanced attention mechanism, which dynamically weights emotional information from text modality and propagates it to acoustic and visual modalities. This mechanism enhances the emotional expression capabilities of non-textual modalities by aligning their features with text-based emotional cues. Second, we utilize the mutual information maximization strategy to strengthen the intermodal associations at the input layer and to refine key task-related information at the fusion layer. This strategy eliminates redundant information and optimizes the representation of multimodal features. Consequently, it thoroughly explores the correlations between different modalities and significantly boosts the effectiveness of modal integration. The experimental results on the CMU-MOSI and CMU-MOSEI datasets validate the effectiveness and practicality of our model in handling complex sentiment analysis challenges.

In this paper, a recurrent spiking neural network model with multiple synaptic connections is constructed, and an online synaptic weight-delay learning algorithm based on spike train kernels is proposed, which can dynamically adjust both synaptic weights and delays during the learning process. A real-time error function is first constructed by applying the kernel function representation of the spike train, and then the online updating rules for synaptic weights and delays are derived by applying the gradient descent method. Spike train learning tasks and nonlinear pattern recognition tasks on UCI datasets are performed to verify the learning performance of the proposed learning algorithm. The experimental results show that the dynamic delay learning algorithm obtains higher learning accuracy in fewer learning epochs than the static delay learning algorithm, and the classification accuracy on UCI datasets is also higher than that of some common supervised learning algorithms for spiking neural networks. It can be seen that the synaptic delay plasticity and the multiple synaptic connection mode can effectively improve the learning performance of recurrent spiking neural networks.

The semantic-guided depth estimation approach can better understand scene information than traditional methods. These methods utilize a shared encoder followed by a two-branch decoder architecture, simplifying the network but failing to capture task-specific details in high-level features. Additionally, some works introduce skip connections or information interaction mechanisms in the decoder to enrich features. However, static feature fusion methods do not adapt well to varying scene changes. To tackle these issues, first, we propose a multi-scale feature refinement mechanism to refine the features of the encoder. Second, we design a dynamic perception fusion decoder that can adjust adaptively during the feature fusion process. Results from experiments show that our approach produces depth maps of excellent quality and continues to perform better in challenging settings.

Pre-trained language models(PLMs) suffer from catastrophic forgetting in continual learning, as sequential task training overwrites previously learned representations. The model’s ability to remain old tasks is referred to as stability, while its adaptability to new tasks is called plasticity. Therefore, the key to addressing this challenge requires balancing model plasticity with stability. To address this issue, in this paper, we propose a novel method to achieve a balance between model stability and plasticity, thereby mitigating catastrophic forgetting. More specific, our proposed approach leverages parameter isolation and subsequent combination strategy. Initially, in training stage, the model adapts on each downstream task via parameter isolation method to prevent potential inference among different tasks. We then combine all trained parameters which containing acquired knowledge by model merging method and finally apply to the backbone model. Empirical evaluations on continual language learning benchmarks substantiate the effectiveness of our approach, revealing a marked enhancement over existing state-of-the-art approaches.

Multivariate time series classification (MTSC) is vital across nume-ous real-world applications, requiring models that effectively capture complex spatio-temporal dependencies within high-dimensional data. Transformer-based architectures have recently established new performance benchmarks in MTSC by effectively modeling long-range dependencies. However, many existing models process spatial and temporal dependencies separately or integrate them only in later stages, thereby limiting their ability to fully exploit spatio-temporal correlations. Moreover, traditional attention mechanisms in Transformer exhibit quadratic computational complexity, hindering scalability and reducing effectiveness for large-scale MTSC tasks. In this paper, we introduce STLANet, a Spatio-Temporal Linear Attention Network, designed to address these challenges by fully integrating spatial and temporal dependencies throughout the model while significantly reducing computational complexity. STLANet incorporates an innovative Time Series Linear Attention (TSLAttn) mechanism that employs BiTransform activation and a local feature enhancement (LFE) approach to enhance feature expressiveness, resulting in more accurate and efficient attention computation. By utilizing distinct spatial and temporal embeddings in conjunction with a cross-attention mechanism, STLANet enables rich local and global spatio-temporal interactions, effectively capturing dependencies across various spatial and temporal levels. Experimental evaluations on 30 UEA MTSC datasets demonstrate that STLANet surpasses state-of-the-art models in classification accuracy with substantially lower computational costs. We believe that STLANet is a robust yet cost-effective approach to high-dimensional, large-scale time series analysis, and it provides a compelling solution for MTSC tasks in resource-constrained environments.

Neural networks have been extensively applied across various domains, where they perform tasks periodically after they are deployed. However, the execution of these tasks often faces changing constraints, such as processing latency and energy consumption. Existing static models, which maintain fixed accuracy and cost, are hard to satisfy these dynamic conditions. To overcome this limitation, we propose a dynamic inference framework based on a dual-degree-of-freedom, multi-route neural network. In this framework, we first design and train the dual-degree-of-freedom multi-route neural network such that there are multiple inference paths with different computational complexity in a single model. Subsequently, we introduce a heuristic routing selection method, which dynamically selects the most appropriate inference routing based on real-time constraints. This approach enhances task accuracy compared to static models and adapting to changing conditions. Experimental results on object detection tasks validate the effectiveness of the proposed dynamic inference framework, demonstrating that our method offers a novel inference strategy for deep neural networks in constrained application scenarios.

In the field of time series data prediction, designing an appropriate neural network requires considering issues such as feature selection and the specific architecture of the network based on the dataset. This demands engineers to have a solid understanding of neural network theory, posing a high entry barrier. Aiming to solve this problem, this paper proposes a two-stage self-adaptive neural network design method called AutoDesign-Net. This method can automatically perform feature selection and construct neural networks without requiring manual intervention in the design details, thus lowering the threshold for engineers. In the first stage, AutoDesign-Net uses an optimized genetic algorithm to automatically select features and determine the optimal neural network architecture suited for the prediction task of the given dataset. In the second stage, based on the network structure obtained in the first stage, the method further optimizes the neural network weights to enhance prediction performance. With this approach, the most suitable feature selection strategy and neural network can be identified for any given dataset, leading to satisfactory prediction results. AutoDesign-Net has been validated on open datasets and demonstrates outstanding performance compared to several existing methods.

With the acceleration of the urbanization process and the increase in population density, the amount of garbage generated is also increasing day by day. All kinds of garbage on the road not only affect the appearance of the city, cause various pollutions, but also affect people’s physical and mental health. The existing algorithms for urban road garbage detection have low detection accuracy and may have omissions and false detections for small-scale garbage. In response to this problem, this paper introduces a SCN-YOLOv8 urban road garbage detection algorithm based on YOLOv8. First of all, in order to better capture the detailed feature information of small objects, we propose a Sub-pixel Spatial Attention Mechanism (SSAM). It can enhance the network’s ability to extract detailed information. Secondly, we have designed the Cross-scale Feature Enhance Pyramid Network (CFEPN). CFEPN not only enhances the interaction ability of feature information in shallow and deep networks, but also enables features with rich contextual information to diffuse to various detection scales. Furthermore, in order to improve the stability of model, we adopt the regression loss function based on the Normalized Wasserstein Distance (NWD) metric. Finally, we constructed a Road Garbage Dataset (RGD), which contains ten common types of road garbage. The experimental results show that our model performs better than the baseline under the RGD dataset, with :0.95 increasing by 2.7%, proving that the modifications we made to the original YOLOv8 algorithm are effective.

Chest X-ray images of patients with pneumonia can quickly and visually show the extent and location of the lung infection, helping doctors to more accurately diagnose and assess the condition. The necessity of segmenting pneumonia X-ray images lies in that it can accurately locate and extract the infected lung area, helping doc-tors to more clearly identify the scope of the lesion and its severity. Therefore, we propose a combined structure U-GANs, which is a pyramid convolutional attention fusion network (PCAF-net) based on Convolutional Networks for Biomedical Image Segmentation (U-net) and a Super-Resolution Generative Adversarial Network (SRW-GAN) based on Generative Adversarial Networks (GAN). This combination not only improves the training quality of the model and enhances the robustness of the diagnostic system, but also provides physicians with more valuable tools for image analysis and assisted diagnosis.

Flight recommendation is to predict the next flight by capturing the passenger travel behavior history. Although existing methods have achieved convincing results in flight recommendation tasks, they ignore the temporal patterns of different periodicity and evolution of passenger preference. In this paper, we propose a novel Multi Hierarchical Time Structures aware Passenger Preference Evolution model for Personalized Flight Recommendation (MTER), which not only uses the time structure of the calendar system to model temporal patterns of passenger multiple behaviors but also attempts to use passenger attributes to alleviate the problems of cold-start and new passengers. Our model divides a passenger behaviors into sessions to form flight-session bipartite graph, and then generates session embeddings by aggregating the corresponding flight embeddings in them, which are further aggregated into the embeddings of different time units. In addition, MTER also captures the passenger attribute embeddings to evolve passenger preference to alleviate the problems of cold-start and new passengers. The latent passenger representations are generated by concatenating all temporal patterns, and attribute embeddings of passengers, which are used to predict the next flight and behavior through a polynomial decoder and a Softmax layer. To the best of our knowledge, this is the first attempt to recommend flight by using multi-hierarchical time structures aware passenger preference evolution model. We conducted experiments on real civil aviation dataset to show the effectiveness of our model.

Fluid simulation is a complex domain that integrates deep learning with traditional methods. This study proposes a novel hybrid fluid simulator that leverages density constraints to model fluid dynamics. By integrating Position-Based Fluids (PBF) with deep learning, our method enhances computational efficiency and ensures stable, physically accurate simulations. Key innovations include: (1) integrating density constraints into neural networks; (2) using spatial hashing to accelerate feature extraction; (3) developing a hybrid simulator that combines PBF stability with neural network efficiency. Experiments show superior performance in visual quality, physical fidelity, and computational efficiency compared to existing methods.

Link prediction plays a crucial role in graph analysis and is widely used in areas like social network analysis, bioinformatics, and recommendation systems. The majority of existing techniques depend on a solitary model, often making it challenging to accurately represent the graph’s overall structure and intricate nonlinear connections, thereby restricting the precision of predictions. In response to this issue, the paper presents an enhanced structure that amalgamates SVD with VGAE. SVD is proficient in capturing the global structure of the graph and can generate low-dimensional representations that retain basic global features. The VGAE learns in the latent space. This is used to model the nonlinear relationship between nodes. In this paper, these two techniques are placed in a unified framework with a joint loss function. Consequently, the model enhances the depiction of information on both a global and local scale. Experimental outcomes on various benchmark datasets indicate the superiority of this model over the standalone SVD and VGAE techniques.

With the advancement of deep learning technology, automatic sleep staging methods based on electroencephalogram (EEG) have garnered widespread attention. Despite previous research on sleep staging achieving high classification performance, several challenges remain unresolved: 1) How to effectively extract and integrate features of sleep at different scales. 2) How to enhance the model’s ability to focus on key features of sleep. 3) How to address the issue of class imbalance in training samples. To tackle these challenges, we incorporate an end-to-end hybrid architecture, CTMASleep, which significantly enhances the model’s ability to extract global and local features by integrating a deep temporal feature extraction module and a contextual temporal fusion module. Additionally, by incorporating a sequence feature reconstruction task, the model’s attention to key feature structures is strengthened. Furthermore, we propose a dynamic class balance loss function to address the class imbalance issue. Experimental results demonstrate that CTMASleep outperforms existing state-of-the-art models on both the Sleep-EDF-20 and Sleep-EDF-78 datasets.

Recent multivariate time series classification methods typically extract features from all variables (channels) but struggle to dynamically analyze and select channels within the model, and usually focus only on the local dependencies between channels. To address the above challenges, we propose a channel scores module based on the self-attention mechanism, which is designed to capture global dependencies between channels and generate channel scores. By adding the channel scores to the time series data, the influence of important channels is effectively enhanced, while the impact of less relevant channels is diminished, thereby boosting the classification performance of the model. We further propose a multivariate time series classification model based on the Transformer architecture, named ChannelFormer, which adds the channel scores module to the model framework to improve the quality and utilization efficiency of time series data. Experiments on 30 benchmark datasets show that our model outperforms other state-of-the-art methods across multiple metrics, demonstrating its significant potential and exhibiting the effectiveness of the proposed channel scores.

Temporal fractional differential equations (FDEs) offer a powerful framework for modeling systems with memory and hereditary effects, commonly encountered in anomalous diffusion, viscoelasticity, and biological processes. However, their intrinsic non-locality presents significant challenges for conventional numerical methods. To address this, we propose a Global-Adaptive Physics-Informed Neural Network (GA-fPINN), an enhanced framework built upon standard PINNs. GA-fPINN integrates a deep neural network for discretizing fractional operators and embeds interpolation-based approximations to improve solution fidelity. Furthermore, it introduces three key adaptive technology: an uncertainty-aware loss function, a dual-parameter adaptive activation function, and a two-branch network architecture. Extensive evaluations on three representative FDE problems demonstrate that GA-fPINN consistently achieves higher accuracy and reduced computational cost compared to the baseline fPINN. These results highlight the framework’s potential for efficient and accurate modeling of complex fractional-order systems.

This study presents a deep reinforcement learning method based on an attention-enhanced spliced heterogeneous graph neural network to address the distributed hybrid flow shop scheduling problem with degradation effect (DHFSP-DE). A spliced heterogeneous graph is designed to represent the complex topology of DHFSP-DE, capturing the interactions among operations, jobs, machines, and factories. To enhance the representation of scheduling states, this work constructs a feature extractor using a graph neural network with a heterogeneous composite attention mechanism that captures complex relationships among different node and edge types. Proximal Policy Optimization (PPO) is used to iteratively refine the scheduling policy and improve action selection. Empirical evaluations on multiple benchmarks show that the proposed model outperforms six widely used dispatching rules in terms of convergence speed and generalization capability.

A comprehensive framework for quantum state estimation in multi-dimensional discrete-time quantum walks is developed, focusing on d-dimensional space. The coin evolution unitary operator is decomposed to characterize the evolution path of the quantum walk, yielding an explicit computational formula for the particle’s position probability distribution. Spectral decomposition of the evolution path establishes rigorous upper and lower bounds for the position probability distribution, demonstrating convergence within a well-defined range regardless of the initial quantum state. Building on these spectral bounds, an inverse problem is addressed: given a target position probability distribution, efficient algorithms are proposed for reconstructing the corresponding quantum initial state. This work overcomes the limitations of conventional forward-simulation methods.

Infertility affects millions of couples globally, with In Vitro Fertilization (IVF) serving as a critical treatment option. The morphological characteristics of human blastocyst components, such as the Inner Cell Mass (ICM) and Trophectoderm (TE) cells, are highly correlated with the success rate of IVF. However, conventional manual assessment of these components is labor-intensive, subjective, and prone to variability. To address these limitations, we propose an advanced deep learning framework, BlastocystMask, designed to automatically segment the internal structures of blastocyst images and enhance objectivity of morphological assessment. BlastocystMask is a two-stage instance segmentation framework. In stage one, it combines Res2Net and Deformable Convolutional Networks (DCN) for robust feature extraction and uses an FPN enhanced with CARAFE for multi-scale feature fusion. In stage two, PointRend refines the segmentation masks by focusing on uncertain regions, effectively addressing cell adhesion and blurred boundaries. Trained on a public human blastocyst dataset with expert annotations, BlastocystMask achieves superior performance, with a Dice coefficient of 91.9% and Jaccard index of 85.0%, outperforming existing methods. Ablation studies confirm that each module contributes to performance gains. BlastocystMask accurately identifies ICM regions and individual TE cells along the blastocyst’s equatorial plane, while introducing quantitative morphological parameters (e.g., cell size, shape uniformity) to complement subjective embryologist by automating segmentation and providing objective morphological metrics.

Traditional intrusion detection systems have certain limitations in dynamic traffic adaptation and real-time processing efficiency. Spiking neural networks (SNNs) have demonstrated unique advantages in temporal information processing. Aiming at the problem that synaptic delays are difficult to be dynamically optimized in the training of traditional SNNs, we propose a delay learning algorithm based on spike train kernels for feedforward SNNs and applies it to network intrusion detection tasks. The proposed algorithm improves the network traffic temporal feature extraction capability of SNNs by co-optimizing the synaptic weights and delays to achieve efficient network intrusion detection. The performance of the proposed algorithm is verified by the NSL-KDD dataset, and compared with the static synaptic delay learning algorithm and other mainstream network intrusion detection algorithms. The experimental results show that the proposed algorithm can effectively improve the network intrusion detection performance of SNNs, achieving detection results comparable to other mainstream network intrusion detection algorithms.

The task of detecting objects in images affected by blurring presents a crucial and difficult problem in the area of object detection. The presence of blurred edges and noise within such images often complicates accurate object localization and identification, making it an issue that demands effective solutions. This study proposes an optimized object detection algorithm that builds upon YOLOv11s, specifically designed to improve performance in detecting objects in blurred images. To tackle feature insufficiency during down sampling and support multi-scale feature extraction, we replace the traditional convolution module with the ADown convolution module. This modification reduces the parameter count, improves model efficiency, and preserves more image information. Additionally, to facilitate the model in focusing more precisely on the blurred regions of images, we integrate the SimAM module to C2PSA. This incorporation significantly improves the precision and reliability of object detection, and at the same time, enhances the computational efficiency. To further optimize model performance, we incorporate the Powerful-IOU loss function, which accords more weightage to arduous samples, encouraging the model to perform better in such cases. Findings from experiments using publicly shared datasets prove that the presented DAP-YOLOv11s model achieves significant improvements across various evaluation metrics. On the vehiclesCounting and NEU-DET datasets, DAP-YOLOv11s performed better than YOLOv11s, with significant improvements in recall, mAP50, and mAP50-95 while reducing the number of parameters.

Neuron coverage quantifies the number and distribution of activated neurons in Deep Neural Networks (DNNs) during testing, serving as a critical metric for evaluating DNN adequacy. It reveals the internal logic of models and uncovers potential issues. Adversarial samples, generated by introducing subtle perturbations to inputs, induce incorrect model predictions and are essential for assessing a model’s robustness against anomalous inputs. Enhancing neuron coverage activates more model pathways, thereby increasing the diversity of test samples. Utilizing coverage feedback to generate adversarial samples optimizes perturbation directions, enhances attack effectiveness, and improves the comprehensiveness and efficacy of testing by exploring diverse model pathways. Here, we present a diffusion-based neuron coverage feedback fuzz testing method, which aims to improve the sufficiency and robustness of DNN testing through a two-stage collaborative optimization framework combining gradient-guided initial perturbation generation and SDEdit-driven naturalness enhancement. This approach integrates gradient optimization with a diffusion generation process to maintain sample naturalness while strengthening adversarial capabilities against target classifiers. During generation, neuron coverage serves as a feedback mechanism guiding the creation and optimization of adversarial samples. Experimental results demonstrate that our method significantly increases both the misclassification rates and neuron coverage across multiple datasets, including MNIST, CIFAR-10, and ImageNet, while maintaining a favorable balance in sample naturalness. Ablation studies further confirm the pivotal roles of the neuron coverage feedback module and the diffusion process in enhancing adversarial sample effectiveness.

Adversarial training (AT) significantly improves adversarial robustness, but generating adversarial examples (AEs) is costly. Fast adversarial training (FAT) reduces the cost, but faces the catastrophic overfitting (CO) problem. Existing solutions adopt a fixed strategy, hindering the model from adapting to robustness differences due to example changes. Thus, an adaptive self-knowledge dynamic guidance FAT algorithm, FGSM-ASKDG, is proposed from the perspective of example changes. First, the adversarial initialization strategy of the gradient momentum perturbation is dynamically adjusted based on the degree of robustness of the model to generate high-quality AEs. Second, the cross-entropy loss of abnormal AEs and the dynamically guided regularization term are introduced to penalize the abnormal AEs and mitigate training instability. Third, the degree of label relaxation is dynamically adjusted in accuracy based on the initialized examples to balance internal and external optimization. Experiments on three datasets and three backbones show that the proposed method effectively prevents CO, remarkably improves adversarial robustness, and is more than three times computationally efficient than the multi-step AT algorithm.

Most privacy-preserving action recognition (PPAR) methods primarily address spatial-domain privacy removal, often neglecting privacy risks in the frequency domain. Additionally, current architectures often struggle to capture fine-grained privacy features and handle high-frequency information. To address these issues, we propose a dual-domain multiscale perturbation framework based on adversarial training, combining frequency-domain perturbation and spatial-domain feature learning. The anonymization module utilizes a Swinv2-Unet architecture, incorporating Swinv2-T as the encoder and U-Net as the decoder. A novel Wavelet Frequency Intervention Module (WFIM) decomposes video frames into high-frequency subbands containing privacy-sensitive details and low-frequency subbands conveying action trends. Learnable Laplacian noise suppresses high-frequency privacy information, while attention mechanisms enhance low-frequency action features, embedding these refined features at multiple decoder levels. Additionally, we design a lightweight cross-domain interaction module that dynamically fuses frequency-domain and spatial-domain features using Neighborhood Attention and its cross-attention variant. Experimental results show that our method achieves stronger privacy protection than previous approaches, with only a slight drop in action recognition performance, demonstrating an effective balance between privacy preservation and task utility.

This paper presents a novel method for recognizing expiration dates that is optimized for the structural complexity and inference speed on food packages, designed to assist vision-impaired individuals. Utilizing computer vision techniques, specifically convolutional neural networks (CNNs), the method efficiently extracts date components (day, month, and year) and converts them into a readable date-time format. Validated on real-world datasets, the proposed method shows significant speed improvements, achieving a precision of 0.9850, a recall of 0.8814, and an F1 score of 0.9303, which is equivalent to the state of the art. The approach demonstrates superior performance compared to existing expiration date recognition systems, ensuring fast inference and autonomy for vision-impaired users.

To achieve full-cycle monitoring of pomegranate growth under resource-constrained scenarios, this study proposes an improved lightweight algorithm YOLO-LFS, based on the YOLOv11 model. Firstly, ShuffleNetv2 is adopted to replace the backbone network of YOLOv11, and the MBConv module is introduced to replace the original detection head, reducing the computational complexity of the model and improving real-time detection capability. The Slide Loss function is incorporated to enhance spatial local information, increasing the focus on hard samples and thereby improving detection accuracy. Secondly, ablation experiments were conducted to compare the performance of different lightweight backbones and loss functions, validating the effectiveness of the modules selected in this study. Finally, the results of the controlled trial indicated that under the same conditions, YOLO-LFS can achieve efficient and accurate detection of pomegranate growth status while maintaining precision. This study provides a lightweight and high-accuracy solution for real-time monitoring and analysis of crops in agricultural production.

Document-level relation extraction serves as a cornerstone for constructing structured knowledge. However, existing methods heavily rely on dense supervision signals and suffer from severe long-tail effects in distantly supervised scenarios characterized by sparse annotations and significant noise, resulting in restricted generalization capability on low-frequency relations and substantial performance degradation. To address these challenges, we propose PaMeRE, a partitioned memory method that enhances long-tail relation extraction through a dual memory mechanism: a general-purpose memory slot captures high-frequency relational patterns, while a dedicated long-tail memory slot focuses on modeling rare relations. The framework employs the Adaptive Fusion Gate(AFG) to achieve context-aware dynamic feature fusion, effectively integrating global semantic patterns with fine-grained relational features. Furthermore, we design a Dynamic Confidence Discrimination Loss that suppresses noisy signals through adaptive focal modulation and margin ranking constraints, prioritizing the identification of genuine relational patterns. Experiments on the ReDocRED benchmark demonstrate that PaMeRE achieves state-of-the-art performance under distant supervision (overall F1 score improved by 3% and long-tail relations F1 score improved by 4%).

Underwater object detection is critically significant in domains such as underwater resource exploration and marine environmental monitoring. However, due to the complex conditions of underwater environment, underwater imaging often suffers from issues such as blurred object boundaries, color distortion, and small object sizes, which limit the performance of traditional general-purpose object detection algorithms, even with the newly developed YOLOv11. To address these challenges, we propose an enhanced underwater small object detection framework based on edge enhancement, termed USEE-YOLO. To tackle the problem of blurred object boundaries in underwater images, we integrate edge enhancement and reassembly modules that reconstruct object boundary information at both the image and feature map levels. To rectify color distortion, USEE-YOLO adopts histogram-based image enhancement methods to optimize color representation. In addition, the framework incorporates a specialized detection head designed for small objects to improve recognition performance. Experiments conducted on the DUO and RUOD datasets show that USEE-YOLO improves $${\text{mAP}}_{50:95}$$ mAP 50 : 95 by 2.2% and 1.5%, respectively, compared to YOLOv11, demonstrating its superior performance and robustness.

Learning-based multi-view stereo (MVS) has advanced significantly in inferring depth maps and reconstructing scenes by matching and fusing images from multiple viewpoints, enabling the acquisition of more comprehensive and accurate 3D information. Although significant progress has been made, certain methods still face challenges such as high computational complexity, loss of high-frequency information, and inaccuracies in feature matching. To solve the above problems, we propose WDSNet model, a wavelet transform and depthwise separable convolution based lightweight MVS algorithm designed for 3D reconstruction. The WDSNet model mainly includes the Wavelet Transform-based Feature Pyramid Network (WTFPN) and the Lightweight 3D Harmonize UNet (LHUNet). In WTFPN, we design a feature pyramid network combined with Wavelet Transform, using low-pass and high-pass filters to alleviate high-frequency information loss, effectively preserving important details while reducing the number of learnable parameters. Then, to reduce memory consumption, LHUNet incorporates 3D Depthwise Separable Convolution (3DDS), which decomposes traditional convolution to optimize computational efficiency while minimizing performance degradation. Additionally, 3D Harmonize Attention (3DHA) module is designed to enhance feature matching accuracy by mitigating matching errors across different depths. Experimental results show that our method not only significantly reduces the memory consumption, but also has advantages over other comparison methods in terms of reconstruction results.

Convolutional neural networks (CNNs) have demonstrated exceptional capabilities across diverse visual recognition tasks. While network pruning serves as a vital technique for compressing overparameterized models effectively, recent investigations reveal its unintended consequences in exacerbating bias among demographic subgroups. Although some research has identified potential factors contributing to unfairness, they lack practical guidance for mitigating it. Driven by the proven efficacy of geometric factors of features (i.e., length and angle) in enhancing robustness of pruned model, we systematically explores their underexplored connections with fairness in pruned networks through comprehensive empirical evaluations. Building upon these insights, we develop FairGEO (Fairness in pruning based on GEOmetric factors of features), a lightweight framework that adjusts the disparities in the geometric factors across subgroups to improve fairness in pruning. Evaluations on multiple image classification benchmarks demonstrate FairGEO can improve fairness while maintaining accuracy compared to pruned baseline and its integration into existing pruning methods confirms its generalizability in enhancing fairness in pruning.

Multilayer Perceptron (MLP), as a simple yet powerful model, continues to be widely used in classification and regression tasks. However, traditional MLPs often struggle to efficiently capture nonlinear relationships in load data when dealing with complex datasets. Kolmogorov-Arnold Networks (KAN), inspired by the Kolmogorov-Arnold representation theorem, have shown promising capabilities in modeling complex nonlinear relationships. In this study, we explore the performance of KANs within federated learning (FL) frameworks and compare them to traditional Multilayer Perceptrons. Our experiments, conducted across four diverse datasets demonstrate that KANs consistently outperform MLPs in terms of accuracy, stability, and convergence efficiency. KANs exhibit remarkable robustness under varying client numbers and nonIID data distributions, maintaining superior performance even as client heterogeneity increases. Notably, KANs require fewer communication rounds to converge compared to MLPs, highlighting their efficiency in FL scenarios. Additionally, we evaluate multiple parameter aggregation strategies, with trimmed mean and FedProx emerging as the most effective for optimizing KAN performance. These findings establish KANs as a robust and scalable alternative to MLPs for federated learning tasks, paving the way for their application in decentralized and privacy preserving environments.

Diffusion models have demonstrated significant performance in generative tasks and have attracted widespread attention. However, the high computational cost of the noise estimation network and the iterative generation process limit their widespread application. Existing caching techniques reduce this burden without additional training, yet error accumulation during generation degrades image quality. To address this issue, we propose a Kalman Filtering Correction method for diffusion model acceleration, which is termed as KFC. In particular, an estimation strategy is designed to mitigate cache-induced errors at each generation step without modifying the architecture of the full diffusion model. It is compatible with existing caching mechanisms and enhances efficiency. Experiments on various diffusion models and benchmark datasets validate its effectiveness.

The rapid growth of Internet of Things devices has intensified cybersecurity threats, necessitating efficient and lightweight intrusion detection systems. Current methods struggle with computational efficiency and feature coverage, particularly for long time-series attack patterns in resource-constrained environments. This study introduces GNN-CST, a novel dual-flow network architecture combining graph neural networks and convolutional neural networks to address these challenges. We developed GNN-CST, integrating a CNN branch to extract local temporal features from traffic data, a GNN branch to model topological interactions among IoT devices, and an adaptive sparse attention mechanism to dynamically focus on critical temporal windows. This design enhances detection of bursty attacks while reducing computational overhead. A cross-modal fusion layer combines local and global features for robust decision-making. GNN-CST exhibits strong performance across six datasets—Edge-IIoT, UNSW-NB15, CICIDS2017, CICIDS2018, BoT-IoT, and ToN-IoT—achieving detection accuracies of 99.46%, 93.50%, 99.84%, 99.78%, 99.99%, and 99.95%, respectively. The average single-epoch training time is 180 s, with a training loss of 0.032, surpassing the baseline Transformer and other models. Compared to the baseline CST model, GNN-CST improves accuracy on multiple datasets and reduces training time by 18%. This study offers an efficient solution for multimodal threat detection in IoT environments.

Anomaly detection in computer vision is typically framed as a one-class classification and segmentation problem. While reconstruction-based methods are widely used, existing networks often suffer from generalization on test data and sensitivity to background noise. We propose a two-stage framework with encoder freezing and hybrid learning, which comprises three key components: (1) a pre-trained encoder freezing to guide normal feature learning, (2) odd-even dataset partitioning for separate reconstruction and discrimination training, and (3) hybrid inputs to enhance noise robustness. Evaluation on industrial inspection benchmarks shows that our method achieves outstanding performance with 98.2% image-level AUC and 75.1% pixel-level AP.

As the complexity of integrated circuit (IC) design increases and its manufacture globalizes, IC designers are increasingly adopting third-party intellectual property (IP) cores to reduce costs and expedite development. However, this shift heightens security risks, particularly the threat of Hardware Trojan (HT). Traditional HT detection methods based on neural network are focused on digital circuits, thus not applicable to analog circuits. To bridge this gap, we introduce GAT-Trans, a Graph Attention Network (GAT) model tailored for detecting analog hardware Trojans, such as A2 Trojan, at the transistor level. Since GAT model enhances weights for critical neighboring nodes, GAT-Trans can improve the detection performance of potential HTs embedded within analog circuits. In the paper, we assess the influence of various node definition methods on detection efficacy and innovate a graph construction method that treats ports of one transistor as a node. We systematically explore different feature extraction methods from analog circuits and evaluate model performance across different feature combinations. This paper details GAT-Trans and proves its effectiveness in identifying A2 Trojan, offering a novel security tool for analog circuits and enhancing defenses against advanced hardware threats.

Clickbait, which aims to induce users with some surprising and even thrilling headlines for increasing click-through rates, permeates almost all online content publishers, such as news portals and social media. Recently, Large Language Models (LLMs) have emerged as a powerful instrument and achieved tremendous success in a series of NLP downstream tasks. However, it is not yet known whether LLMs can be served as a high-quality clickbait detection system. In this paper, we analyze the performance of LLMs in the few-shot and zero-shot scenarios on several English and Chinese benchmark datasets. Experimental results show that LLMs cannot achieve the best results compared to the state-of-the-art deep and fine-tuning PLMs methods. Different from human intuition, the experiments demonstrated that LLMs cannot make satisfied clickbait detection just by the headlines.

Trajectory control for stratospheric balloons relies on accurate modeling of complex wind fields.However, stratospheric winds exhibit complex multi-scale dynamics and spatio-temporal non-stationarity, posing significant modeling challenges. While existing high-fidelity numerical models are computationally expensive and unsuitable for developing control strategies like reinforcement learning (RL), purely data-driven models often lack physical realism. We propose a Wavelet-Fourier Physics-informed Latent Diffusion Model (WF-PLDM). It combines Wavelet and Fourier transforms to capture the wind field’s multi-scale characteristics and integrates Navier-Stokes equations to ensure physical consistency. Quantitative evaluations on the ERA5 reanalysis dataset demonstrate that the proposed WF-PLDM outperforms several baseline generative models in generation accuracy and physical consistency metrics. Our approach provides a reliable simulation environment conducive to training advanced RL-based control strategies for stratospheric balloons.

The Cryptographic Algorithm Recognition (CAR) task is a critical problem in cryptography, with significant implications for the security of cryptographic algorithm design. While Large Language Models (LLMs) demonstrate promising potential in addressing this task, evaluating their performance remains a challenge due to the absence of aligned input-output specifications and standardized evaluation metrics for it in LLMs. In this paper, we construct an evaluation dataset and the corresponding metrics to analyze the performance and factors that influence effectiveness in CAR. The evaluation includes seven different cryptographic algorithms, along with performance of five main LLMs in this dataset. Experimental results indicate that LLMs exhibit limitations in algorithm identification, achieving an average accuracy of 63.9%. The performance is significantly influenced by the cryptographic algorithm and the fundamental capabilities of LLMs. Surprisingly, a mainstream cryptographic algorithm called Keccak can be relatively recognized by LLMs, which it shouldn’t be, unlike other modern algorithms. Furthermore, we introduce six different prompt engineering methods and find that most do not significantly enhance LLM performance in CAR. However, the prompting approach of snapshot-based exemplar reference effectively improves performance of CAR, resulting in an average increase of 7.7%, with varying degrees of improvement under different conditions.

Subject-oriented video captioning generates natural language descriptions focusing on a specified subject’s activities. While prior works mainly utilize subject input to guide frame extraction and content encoding, they often neglect subject-environment interactions and detailed action modeling, leading to inaccuracies in action-related captions. To tackle this issue, we propose an enhanced framework that leverages predicate constraints and hierarchical interaction mechanisms to guide the model in understanding the actions of the subjects in the video, thereby enabling a more accurate summary of the subjects’ behaviors and generating more precise captions. Specifically, we first adopt predicates extracted from captions as supervision signal for learning subject-related action feature. Next, we utilize Swin-Transformer to extract hierarchical video features and facilitate hierarchical interactions between the subject features and the output of each layer, thereby capturing action semantics across various scales. To refine action feature learning, we introduce a composite loss function combining cosine similarity and mean squared error (MSE) to ensure alignment with the target embeddings. Finally, a gate mechanism dynamically fuses the video, subject, and action features, which are then input into a BERT-based generator to produce high-quality captions. Experimental results on the SO-MSVD and SO-MSRVTT datasets demonstrate that our approach significantly enhances captioning accuracy, particularly in terms of action verb precision, while maintaining computational efficiency suitable for real-world applications.

In surveillance, subjects are often out of camera range, leading to low-resolution, unrecognizable face images. Addressing this, we propose a deep learning approach for single face super-resolution (SR) that focuses on enhancing facial features by weakening non-primary information with a novel loss function (WNI-L). This function prioritizes facial clarity over background, improving recognition. Additionally, we introduce an activation function with a threshold (ReLU-T) to normalize brightness variations, crucial for SR. Our method, combining WNI-L and ReLU-T, outperforms existing SOTA methods.

To enhance the operational efficiency of power grids and mitigate losses, this study proposes an abnormal electricity consumption detection method grounded in the Time Series Generative Adversarial Network (Time GAN) and the CNN-Transformer-Bi LSTM architecture. Initially, the raw dataset is preprocessed through data cleansing and missing value imputation. Subsequently, to address the scarcity and low prevalence of abnormal samples in real-world data, a Time-GAN data augmentation approach integrated with a multi-head self-attention mechanism is put forward. The incorporation of this mechanism effectively elevates the quality of generated abnormal data. Next, a hybrid CNN-Transformer-Bi LSTM model is utilized. By reducing the dimensionality of CNN kernels, it extracts local short-term periodicities in sequential data. Transformer is designed to capture long-range global dependencies, and Bi LSTM is employed to extract time-dependent features of column-based data. Their synergy enhances the model’s capacity to learn sample characteristics. Validation on real-world datasets demonstrates that the proposed method outperforms existing models in comprehensive metrics such as accuracy (ACC) and the area under the curve (AUC) value.