Abstract:With the rapid development of artificial intelligence and intelligent vision technologies， images and videos have evolved from media primarily intended for human viewing into essential sources of information for machine perception， understanding， and decision-making. Conventional visual coding methods primarily optimize pixel reconstruction fidelity and perceptual quality for human vision. However， these objectives do not necessarily ensure the efficient preservation of semantic information and discriminative features required by machine vision tasks， such as object detection， semantic segmentation， object tracking， and action recognition. Coding for machine vision has therefore attracted increasing attention from both academia and industry as an important research direction in visual information compression. This paper systematically reviews its key technologies and development trends， covering the optimization of conventional coding standards and deep-learning-based coding methods. First， machine-oriented optimization methods for conventional image and video coding standards are examined， with particular attention to task-aware bit allocation， rate-distortion optimization for machine vision， and coding tool optimization. Second， deep-learning-based methods are organized into three categories according to the coding object： Learned image coding， learned video coding， and learned feature coding. The evolution of these technologies from pixel reconstruction toward semantic preservation， task performance preservation， and edge-cloud collaborative intelligent processing is analyzed. Finally， the paper discusses major challenges concerning evaluation protocols， generalization across tasks and models， computational complexity， compatibility with existing standards， and the joint requirements of human-machine collaboration and privacy protection. Future research directions are also outlined， including unified rate-distortion modeling， general-purpose visual representation coding， lightweight intelligent coding， human-machine collaborative coding， and standardization.Highlights:1.A systematic review of coding for machine vision is presented， covering conventional coding standard optimization and learned image， video， and feature coding.2.Representative methods are compared in terms of core ideas， key techniques， and coding performance， tracing the shift from pixel reconstruction to semantic preservation and task-oriented compression.3.Major challenges and future directions are discussed， including unified rate-distortion modeling， general-purpose visual representation coding， lightweight deployment， human-machine collaborative coding， and standardization.

Abstract:Image forgery detection and localization （IFDL） determines whether an image has been manipulated or synthesized in its entirety and identifies locally manipulated regions. Generative editing and whole-image synthesis change the available forensic traces and broaden the scope of detection research. This survey reviews representative literature from 2005 to 2026. Conventional local manipulation， generative local editing and fully synthetic images are distinguished by how their pixels are produced. Their traces， datasets and evaluation metrics are summarized. Representative methods are then discussed along three research directions： conventional forensic cues， generative forgery analysis and multimodal large language models. The review compares their mechanisms， applicable conditions， evaluation settings and performance under transmission degradation. The analysis shows that image-level labels and pixel-level masks support different evaluation tasks， while the reported composition of genuine and forged images is inconsistent across publications for some datasets. Training data， supervision， input resolution and threshold selection affect performance comparisons； some differences in evaluation settings also change method rankings. Under generative editing， classical cues depend on background preservation， region blending and subsequent processing. Generator artifacts， diffusion-inversion residuals and pretrained features provide complementary evidence， but generalization remains sensitive to differences between generators and data distributions. Multimodal large language models introduce natural-language explanations， yet methods differ in their support for localization masks and source attribution. Explanation readability and faithfulness to visual evidence require separate assessment. The effects of compression， resizing and noise are related to the cues used； controlled experiments are still needed to distinguish the contributions of training data and network architecture. This survey provides a reference for selecting forensic evidence， designing comparable evaluations and choosing methods for practical deployment.

Abstract:Modern video coding standards such as the audio video coding standard 3 （AVS3） have substantially improved compression efficiency， but conventional rate-distortion optimization still relies heavily on signal-fidelity measures such as mean squared error （MSE） and peak signal-to-noise ratio （PSNR）， which do not fully reflect the nonlinear characteristics of the human visual system （HVS）. Knowledge-distillation-based perceptual coding methods can improve visual quality by transferring bit-allocation knowledge from perceptually optimized learned image compression models to block-level quantization parameter （QP） maps. However， most existing approaches are designed and evaluated under image or all-intra coding configurations. When directly extended to practical video coding with both intra- and inter-frame prediction， a fixed perceptual QP adjustment strategy may produce content-dependent performance fluctuations， especially for sequences with complex textures or strong temporal variation. To address this problem， this paper proposes a lightweight content-adaptive QP decision method for AVS3 video coding. Before actual encoding， a video content pre-analysis module processes the first ten frames of each segment using a pretrained VGG-16 network. Features from shallow， intermediate， and deep layers are pooled and fused to describe fine-grained texture， local structure， and high-level semantics. Temporal variation of the fused features is further summarized into four compact descriptors： Texture， motion， complexity， and motion consistency. A logistic-regression classifier then generates a binary gating signal to determine whether the downstream perceptual optimization module is enabled or bypassed. The classifier is trained on 292 non-overlapping 32-frame segments constructed from standard HEVC test sequences， where the ground-truth label of each segment is determined by the MS-SSIM BD-rate difference between perceptual optimization and the AVS3 anchor under identical coding configurations. Five-fold stratified cross-validation yields an accuracy of 92.3%， a precision of 82.0%， a recall of 93.3%， and an F ₁ score of 87.3%. Experiments on representative 4K UHD sequences under random-access configuration show that， relative to the AVS3 HPM 12.0 anchor， the proposed method reduces the average BD-rate by 7.11% under multi-scale structural similarity （MS-SSIM） and by 3.39% under PSNR. When integrated into a highly optimized content-aware code （CAC）， the average MS-SSIM BD-rate relative to the common HPM anchor improves from -21.45% to -26.24%， corresponding to an additional gain of 4.79 percentage points， while the average PSNR BD-rate improves from -3.60% to -7.01%. On a synthetic 64-frame sequence containing abrupt scene changes， the proposed adaptive scheme achieves an 11.82% MS-SSIM BD-rate reduction and a -0.53% PSNR BD-rate， whereas unconditional perceptual distillation yields a 6.79% MS-SSIM reduction but a 1.71% PSNR increase. These results indicate that the proposed segment-level gating mechanism can retain perceptual coding gains while reducing the risk of objective-quality degradation caused by unconditional perceptual adjustment. The method does not alter the standard bitstream structure and can be integrated as a low-overhead encoder-side decision module in conventional hybrid video coding frameworks.

Abstract:Low-altitude air-ground communication channels are characterized by time-varying multipath propagation and dynamic switching of line-of-sight （LOS） and non-line-of-sight （NLOS） components. Traditional pure data-driven joint source-channel coding methods lack physical communication priors， leading to poor image reconstruction robustness and oscillating training convergence. This paper proposes an end-to-end image semantic communication architecture （UP-WITT） with model-data dual-driven design for low-altitude scenarios. Taking the time-domain baseband transmission model of low-altitude time-varying channels as the physical basis， the architecture adopts a two-branch collaborative structure. The data-driven branch employs Swin Transformer as the codec backbone to extract multi-scale global semantic features， which are further mapped into time-domain discrete baseband symbol sequences complying with transmission constraints. Based on projected gradient descent （PGD） theory， the model-driven branch unrolls the iterative optimization process into learnable network layers and constructs a configurable PGD reconstruction module. With the physical prior of time-domain channel transmission， the module performs effective compensation of channel impairments and reduces iterative complexity via frequency-domain fast computation， thereby alleviating the ill-posedness of the inverse problem under the experimental conditions considered. Experimental results on the CIFAR dataset show that compared with pure data-driven baseline models， the proposed method achieves better peak signal-to-noise ratio （PSNR） and multi-scale structural similarity （MS-SSIM） under different channel bandwidth ratios （CBRs） and channel conditions. It also exhibits a smoother training convergence process and better reconstruction performance within the trained channel distribution， providing a useful reference for reliable image semantic transmission in low-altitude applications.

Abstract:In UAV-based photovoltaic panel defect detection， infrared and visible images provide complementary information on thermal anomalies and surface texture structures. However， local modality misalignment， attenuation of weak responses from tiny defects， and the high computational cost of dual-modal models still limit detection performance and practical deployment. To address these challenges， this study proposes a feature-enhanced lightweight dual-modal detector for UAV-based photovoltaic panel defect detection， termed FLD-DET. Built on an infrared-visible dual-modal detection framework， FLD-DET incorporates a dynamic sparse window （DSWin） module into the backbone feature extraction process. DSWin enhances the retention of weak responses from tiny defects through local convolutional inductive bias， channel compression， and adaptive attention allocation. To mitigate local spatial discrepancies between dual-modal features， a lightweight cross-modal attention （LCoA） mechanism is designed to strengthen the spatial correspondence between infrared and visible features by encoding directional positional information. During the detection feature generation stage， FLD-DET introduces a regional-aware feature aggregation （RAFA） head to enhance the discriminative representation between defect regions and background regions. Experimental results show that FLD-DET achieves a favorable balance between detection performance and inference efficiency on the aerial dual-modal photovoltaic panel defect dataset and demonstrates good recognition capability for tiny defects. FLD-DET also obtains competitive detection results on the M3FD dataset， indicating potential cross-scenario generalization.

Abstract:Object detection is a prerequisite task for smart ports to realize behavior recognition of passengers. To address the problems that occlusion caused by crowded queues and gate equipment in fast-track channels leads to decreased detection accuracy of existing models， while also considering the requirements of edge devices， an improved YOLOv12n-based occluded object detection algorithm， MSS-YOLO， is proposed. First， an adaptive multi-scale feature extraction （AMFE） module is designed in the backbone network to replace the downsampling module， enlarge the receptive field of the model， and adaptively weight and focus on key human body feature regions， thereby enhancing multi-scale feature extraction capability and alleviating feature information loss and occlusion interference. Second， an improved feature enhancement module， A2C2f_SEAM， is introduced into the neck network to learn the relationship between occluded and non-occluded regions and dynamically adjust the focus of attention. Finally， CIoU is replaced with Alpha-Shape IoU， and power transformation and geometric constraints are introduced to further optimize bounding box regression performance. Experiments are conducted on a CrowdHuman experimental subset and the self-collected port fast-track channel crossing behavior dataset CBCD. The results show that， compared with the baseline model， the proposed MSS-YOLO improves precision by 1.4 and 6.1 percentage points， recall by 7.1 and 4.3 percentage points， mAP@0.5 by 5.4 and 5.0 percentage points， and mAP@0.5：0.95 by 4.5 and 12.7 percentage points on the two datasets， respectively. Meanwhile， the number of model parameters is reduced by 0.23×10?. MSS-YOLO improves the detection accuracy of occluded persons in port clearance scenarios while maintaining a relatively small parameter scale， providing a foundation for subsequent edge deployment and optimization.

Abstract:As a core enabling technology for trustworthy artificial intelligence， research on the interpretability of machine learning models plays a vital role in automated decision-making systems. Among them， saliency-based explanations， which intuitively reveal the basis of model decisions in the form of heatmaps， have become a key research direction in explainable artificial intelligence. However， existing approaches suffer from fundamental limitations： Perturbation-based methods are constrained by manually designed distributional assumptions， making it difficult to balance computational efficiency and generalization ability； gradient-based methods require prior knowledge of the model’s internal structure and thus fail to explain black-box models； and linear-approximation-based methods introduce model bias due to excessive local simplification. To address these limitations， we propose Polixplain， a policy-gradient-based saliency explanation method for black-box models. By using the Deletion Score as the core reward signal and incorporating a differentiable relaxation technique into the policy gradient optimization process， Polixplain enables end-to-end optimization toward the faithfulness objective. Image classification tasks are adopted as the experimental testbed， where the high-dimensional feature space and intuitive visual representation provide an ideal environment for multidimensional evaluation of explanation methods. Polixplain outperforms a range of existing explanation methods in terms of both explanation faithfulness and the human interpretability of the generated heatmaps.

Abstract:This paper introduces a two-stage deep learning approach to enhance the performance of single-channel target speech extraction. Conventional discriminative models often over-optimize objective metrics， leading to unnatural artifacts， distortions， and compromised subjective auditory quality in the extracted speech. To address this limitation， we innovatively incorporate a diffusion model as a second-stage refinement module. This module， built upon a stochastic differential equation （SDE） framework， regenerates and refines the initial output from the discriminative model. This process not only reduces extraction errors but also significantly improves the naturalness and harmonic clarity of the target speech. Experimental results on the WSJ0-2mix-extr dataset show that our method achieves an 11.18% improvement （from 3.22 to 3.58） on the non-intrusive speech quality assessment （NISQA） metric， which simulates subjective auditory perception. This indicates a substantial enhancement in perceived speech quality and naturalness. Furthermore， subjective listening tests using the comparative mean opinion score （CMOS） confirm the effectiveness of our approach in boosting speech clarity and intelligibility.

Abstract:The resting state functional magnetic resonance imaging （rs-fMRI） technology is widely used for predicting and diagnosing brain diseases. Researchers have proposed a large number of diagnostic methods based on rs-fMRI functional connectivity networks， but these methods have problems such as not fully considering the complex dependency relationships of brain network data and high computational costs. To address the aforementioned issues， this paper proposes a brain disease diagnosis method based on the bi-directional selective state space model （SSM） framework， named as Bi-BrainSSM. This method designs an attentive bi-Mamba （ABM） module that combines multi-head attention mechanism with a bi-directional selective state space model to enhance the representational ability of brain functional connectivity networks and improve the computational efficiency of the model. It simultaneously introduces orthogonal clustering readout to solve the problem of information loss in traditional aggregation methods. This method is validated on the publicly available ABIDE autism dataset and REST-MDD depression dataset. Experimental results show that compared to other comparison methods based on rs-fMRI， Bi-BrainSSM achieves more accurate diagnostic performance， while also having higher computational efficiency and lower GPU memory occupancy.

Abstract:For extremely large-scale multiple-input multiple-output （XL-MIMO） enabled 6G communication systems， the near-field effect poses performance limitations for traditional channel estimation methods based on far-field plane wavefront models. Existing near-field schemes predominantly rely on the Fresnel approximation， which introduces modeling errors， and generally handle channel estimation and multi-user activation detection in separate steps， leading to error accumulation. To address these challenges， this paper proposes a joint near-field channel estimation and multi-user activation detection framework based on an accurate spherical wave channel model. Different from approximation-based near-field schemes， the proposed framework directly characterizes the near-field array manifold without relying on Fresnel approximation， thereby avoiding model mismatch. Furthermore， the proposed method jointly exploits the double sparsity of XL-MIMO channels in both the polar domain and the user domain. Specifically， the polar domain sparsity arises from the limited number of effective propagation paths described by angle and distance parameters， while the user domain sparsity comes from the fact that only a small subset of users is active during a given transmission interval. By leveraging these two types of sparsity simultaneously， the original high-dimensional coupled estimation problem can be transformed into a more structured sparse recovery problem. Moreover， an independent-vector variational Bayesian inference （VBI） technique is introduced to efficiently decouple the complicated joint estimation task into multiple user-wise sparse recovery subproblems. This strategy enables effective separation between the near-field array manifold and multi-user pilot signals， thereby substantially reducing computational complexity while preserving estimation accuracy. In addition， an off-grid parameter update strategy is incorporated to refine the predefined angular and distance grids. By iteratively updating the off-grid parameters， the proposed method alleviates grid mismatch and improves the accuracy of near-field channel reconstruction. Simulation results demonstrate that the proposed method outperforms existing methods in terms of channel estimation accuracy and user activity detection reliability. Under different signal-to-noise ratios， antenna numbers， snapshot lengths， and active user numbers， the proposed algorithm consistently obtains lower normalized mean square error and higher user activity detection success rate. These results verify the effectiveness， robustness， and superiority of the proposed framework. Therefore， the proposed method provides an efficient and accurate solution for near-field XL-MIMO channel acquisition and grant-free massive access in future 6G wireless communication systems.

Abstract:Brain-computer interfaces （BCIs） based on steady-state visual evoked potential （SSVEP） are rapidly advancing in human-computer interaction systems. However， the classification of SSVEP signals in short time windows still faces challenges such as low accuracy and insufficient feature extraction. In this paper， we propose an attention enhancement-dual channel multi-feature convolutional neural network （AE-dCNN）. The network applies a channel attention mechanism to weight the features of different channels， enhancing the representation of useful information. Then， two parallel channels are employed to extract time and frequency domain features from the signals， and the extracted features are fused for classification. Cross-subject and subject-independent experiments are conducted on both public and self-built datasets. The results demonstrate that the proposed AE-dCNN model achieves the highest accuracies of 94.38% in cross-subject experiments among five algorithms and of 92.36% in subject-independent experiments among six algorithms. Additionally， we explore the application of the Kolmogorov-Arnold network （KAN） structure in EEG signal processing. The results indicate that the KAN model outperforms the multilayer perceptron （MLP） model in terms of accuracy across most time windows.

Abstract:To address the challenge of accurately and efficiently detecting cells and injection needle tips with large scale discrepancies in automated microinjection， this paper proposes a high-efficiency instance segmentation model named DELD-YOLO. Built upon the YOLO11 architecture， the proposed model incorporates structurally re-parameterized diverse branch blocks into the backbone to enhance feature representation. An efficient lightweight multi-scale feature pyramid network is introduced to improve cross-scale feature fusion， and a detail-enhanced lightweight segmentation head is designed to preserve fine structural information while reducing computational cost. Experimental results demonstrate that DELD-YOLO outperforms the baseline YOLO11 in both cell and injector segmentation tasks， achieving mAP values of 98.7% for cells and 75.2% for injectors at IoU thresholds from 0.5 to 0.95. With only 2.08×10⁶ parameters and 10.0×10⁹ FLOPs， the proposed model exhibits a favorable balance between accuracy and efficiency， making it well suited for automated cell injection and edge-device deployment.

Abstract:Due to factors such as medical ethics， recording conditions， participant cooperation， and annotation costs， high-quality Parkinson’s disease （PD） speech annotation data is relatively scarce， which to some extent restricts the application of supervised learning methods in speech-based PD detection. Although the above problems can be overcome through cross-lingual detection and the use of multi-types of corpora， most existing studies have adopted homogeneous corpus alignment strategies， ignoring the complexity and diversity of speech expression in natural contexts， making it difficult to fully extract the deep pathological features of PD. To this end， a continuous self-supervised cross-lingual PD detection method is proposed in this paper， aiming to fully utilize speech data resources for PD detection while avoiding time-consuming and laborious manual annotation. First， based on multi-type speech data in the source language， a sequential fine-tuning strategy is adopted to adaptively train the top-level of the self-supervised feature extraction model. To prevent corpus conflicts and forgetting issues， a continuous learning replay buffer and feature distillation mechanism are introduced to enhance the plasticity of the self-supervised feature extraction model while improving its ability to retain old knowledge. Finally， the model is transferred to the target language and the ability of self-supervised models to model “language independent” features is utilized to achieve effective recognition of PD in cross-lingual scenarios. Experimental results on Italian and Chinese PD datasets show that the proposed method extracts more discriminative features and has better classification accuracy and generalization ability than existing methods in cross-lingual PD detection.Highlights1. A continual self-supervised framework is proposed for cross-lingual Parkinson’s disease detection using multi type speech corpora.2. K-means replay buffer and cosine-similarity-based feature distillation mitigate catastrophic forgetting across diverse speech types.3. The HuBERT-based model extracts language-independent pathological features and achieves superior cross-lingual detection performance on Chinese and Italian datasets.

Abstract:For high altitude platform （HAP） communication systems， this paper analyzes the wireless transmission performance under an integrated scheme of space division multiple access （SDMA） and non-orthogonal multiple access （NOMA） based on stochastic network calculus （SNC） theory. First， under the condition that only user location information is available， a user clustering algorithm based on the K-means clustering algorithm is proposed. Second， in scenarios where the HAP is equipped with a uniform linear array， a hybrid multiple access transmission scheme is proposed in which inter-cluster users employ SDMA， while intra-cluster users adopt power-domain NOMA. Furthermore， considering that the HAP-to-user link follows a Nakagami-m distribution， closed-form expressions for the upper bound of the system delay violation probability and effective capacity are derived based on SNC theory for both successive interference cancellation （SIC） and joint decoding （JD） schemes. Finally， computer simulations verify the correctness of the theoretical analysis and reveal that the JD decoding scheme is superior to SIC in reducing system delay and enhancing effective capacity.

Abstract:Deep learning technology has been widely applied to the field of medical image segmentation； however， accurate segmentation of bladder images remains challenging. To address the segmentation of bladder magnetic resonance imaging （MRI） iamges， this study proposes MIV-Net， an improved algorithm based on feature fusion and attention mechanisms. The network introduces a multi-scale core feature fusionizer （MCF） to aggregate feature information at different scales and enhance the representation of local structures and boundary details. A vision Transformer （VIT） network is incorporated to capture long-range dependencies， thereby improving the modeling of global contextual information. In the decoding stage， a multilevel attention fusion （MAF） module is employed to integrate multi-level semantic and spatial information， allowing the network to make better use of contextual cues and recover fine segmentation details. Experiments are conducted on two tasks： Bladder wall segmentation and bladder tumor segmentation. For bladder wall segmentation， MIV-Net achieves IoU， Precision， and Dice coefficient values of 83.16%， 91.66%， and 90.81%， respectively. For bladder tumor segmentation， the corresponding values are 82.49%， 90.50%， and 90.40%. These results demonstrate that MIV-Net outperforms current mainstream methods across the key evaluation metrics and provides accurate segmentation of both bladder walls and tumors. By jointly exploiting multi-scale local features， global dependencies， and multi-level contextual information， the proposed model offers a promising technical approach for automated bladder MRI image analysis and may support more accurate computer-aided diagnosis of bladder cancer.Highlights:1. Propose MIV-Net by integrating the multi-scale core feature fusionizer （MCF）， vision Transformer （VIT） network， and multi-level attention fusion （MAF） module for bladder MRI image segmentation.2. Enhance multi-scale feature extraction， global information modeling， and contextual information fusion through the coordinated use of MCF， VIT， and MAF.3. Achieve Dice scores of 90.81% and 90.40% for bladder wall and bladder tumor segmentations， respectively， outperforming current mainstream methods.

Abstract:To address the challenges of low positioning accuracy and poor robustness caused by the strong nonlinear coupling of piezoelectric hysteresis， mechanical friction， and transmission backlash in the macro-micro composite-driven cell injection mechanism， an adaptive fast non-singular terminal sliding mode （FNTSM） control strategy is proposed. This strategy aims to eliminate the dependence on precise dynamic models and achieve cross-scale precise motion control. First， a dynamic model of the piezoelectric-driven cell injection mechanism is established， considering uncertainties such as hysteresis， friction， and other perturbations. Then， to overcome the dependence on the prior knowledge of disturbance bounds and the chattering issue in traditional sliding mode control， a FNTSM controller is designed that integrates time delay estimation （TDE） technology with adaptive gain adjustment. This controller uses the TDE to estimate and compensate for lumped disturbances in real-time， ensuring system states converge within a finite time through FNTSM. Additionally， a PID-type sliding mode surface is introduced， and an adaptive law is designed to dynamically adjust the sliding mode surface parameters， enhancing the system’s ability to suppress strong time-varying disturbances such as friction coefficient changes and load mutations， as well as improving response speed. Experimental results show that in tracking a sine wave trajectory， the controller’s root mean square error percentage is 1.39%， with a maximum tracking error percentage of 4.79%. Compared with the traditional PIDSM controller， the maximum tracking error is reduced by 50.9%. In comparison with the recently proposed PIDSMC-RBF advanced controller， the maximum tracking error percentage is reduced by 9.28%. The proposed controller effectively improves the trajectory tracking accuracy of the cell injection mechanism under cross-scale motion， providing a reliable technological solution for micro-manipulation operations.

Abstract:Conventional density-based spatial clustering of applications with noise （DBSCAN） algorithms struggle in complex dynamic traffic due to non-uniform density of millimeter-wave radar point clouds， causing ambiguous boundaries， poor separability， and noise misclassification. This paper proposes a two-stage enhanced DBSCAN framework integrating adaptive polar grid partitioning with velocity-based post-processing. In stage one， a dynamic elliptical neighborhood search adapts its range according to radial distance， ensuring accurate initial spatial clustering under density variations. Stage two exploits Doppler radial velocity statistics： Outliers are removed via a velocity-consistency criterion， misclassified valid points are reassigned through a spatial-velocity dual-similarity framework， and overlapping clusters are split based on velocity dispersion and motion direction. Experimental results demonstrate significant improvements over traditional DBSCAN. In vehicle-overlapping scenarios， accuracy rises from 79.59% to 88.1%； in pedestrian-vehicle overlaps， from 86.55% to 97.24%. Evaluation metrics show the silhouette coefficient improves by 14.4%， the Calinski-Harabasz index by 381.3%， and the Davies-Bouldin index decreases by 41.5%， indicating clearer boundaries and better cluster stability. The proposed method effectively resolves overlaps and suppresses noise， offering a reliable preprocessing solution for millimeter-wave radar in dynamic environments.Highlights:1.Adaptive polar-grid DBSCAN with dynamic elliptical neighborhoods compensates for radial-distance-dependent density distortion, ensuring robust initial clustering under non-uniform point-cloud distribution.2.Velocity-consistency filtering and spatial-velocity dual-similarity reassignment accurately remove outliers and recover misclassified valid points, preserving target integrity.3.Overcome the inherent limitation of spatial-only DBSCAN by leveraging velocity dispersion and directional differences to split overlapping clusters, enabling accurate separation of adjacent targets with distinct motions.

Abstract:Addressing the increasingly pronounced near-field effects arising from the development of ultra-large antenna arrays in 6G communications， this paper investigates a covert communication system employing multi-input single-output non-orthogonal multiple access （MISO-NOMA） assisted by a pinching antenna system （PASS）. This research aims to overcome the limitations of conventional flexible antennas in complex wireless environments， specifically addressing deficiencies in signal coverage capability and deployment flexibility. By innovatively deploying positionally adjustable pinching antennas （PAs） on dielectric waveguides， the system effectively harnesses the beam-focusing properties of near-field spherical wave models. An efficient alternating optimization （AO） algorithm is proposed for the joint optimization of transmit power allocation， beamforming vectors and PA positioning. This is achieved by deriving a closed-form power solution， introducing penalty terms to address the rank-one constraint of beamforming， and employing a particle swarm optimization algorithm to locate the optimal PA position. Numerical simulations demonstrate that the proposed scheme significantly enhances system performance in the 28 GHz millimeter-wave band， overcoming the 45 bit/（Hz·s） bottleneck for covert communication rates. Compared with conventional schemes， it achieves a prominent performance improvement， validating the effectiveness and superiority of PASS technology in near-field covert communication.

Abstract:The electroencephalogram （EEG）-based emotion recognition is a key technology in next-generation human-computer interaction and brain-computer interface （BCI） systems， with significant applications in national brain science initiatives and intelligent healthcare. However， the inherent non-stationarity， low signal-to-noise ratio， and scarcity of medical EEG samples present substantial challenges for existing deep learning methods in real-world scenarios. Most current approaches directly extract frequency-domain features from single-channel raw signals. This strategy not only risks misinterpreting high-frequency noise as informative features but also neglects the spatial synergy among different brain regions， limiting the model’s generalization ability. To address these issues， this paper proposes a synergy-first end-to-end EEG emotion recognition framework， introducing the synergy-frequency aggregation network （SFAN） for collaborative frequency-domain feature extraction based on aggregation convolutions. Departing from the conventional “frequency extraction followed by spatial fusion” paradigm， the SFAN first employs a channel aggregation convolution module to adaptively fuse multi-channel EEG signals， simulating dynamic inter-regional brain synergies to generate highly robust collaborative representations. Frequency-domain feature extraction and deep spatiotemporal modeling are then performed on these representations. This strategy effectively suppresses channel-level noise prior to frequency analysis and enriches the feature space through diverse aggregation patterns， alleviating overfitting under small-sample conditions. The SFAN achieves a classification accuracy of 97.42% on the SEED-Ⅳ dataset and 94.91% on the SEED-Ⅴ dataset. Experimental results demonstrate that the proposed method consistently outperforms existing approaches in cross-subject and cross-session emotion recognition tasks， validating the effectiveness of incorporating inter-regional brain collaboration prior to frequency-domain analysis for enhancing EEG emotion recognition robustness.

Abstract:As an embodiment of the core 6G vision of scenario-based mobile communication， the 5G kite private network is designed to address practical challenges in the power industry such as high-frequency terminal mobility， dynamic topological reconfiguration， and low-latency service guarantees. To tackle the issues of insufficient accuracy and weak adaptability in security situation awareness within highly mobile and elastic topological environments of 5G kite private networks， this paper proposes a spatiotemporal dual-modal based elastic security situation awareness method for power networks. The approach constructs a multi-source input vector by collecting core feature data， employs long short-term memory （LSTM） networks for temporal anomaly detection， and integrates isolation forest for spatial anomaly detection， thereby generating an enhanced feature matrix. Furthermore， a bisecting K-means clustering algorithm and a dynamic weight adjustment mechanism are introduced to adaptively optimize feature weights， enabling accurate classification of security situations into three levels： Stable， vulnerable， and threatened. Finally， the weighted principal component analysis is applied for dimensionality reduction and visualization. Simulation results demonstrate that the proposed method effectively responds to changes in network state， improves situational judgment accuracy across various scenarios， and provides reliable support for the security maintenance of 5G kite private networks.