ABSTRACT. This paper presents a novel binary ant lion optimizer (NBAL) to solve 0-1 knapsack problem. In the algorithm, NBAL uses two tuples, which consists of real-valued vectors and binary vectors, to represent the antlion and ant individuals. Therefore, NBAL acts directly on real-valued vectors, while the solutions are represented by binary vectors, thus ensuring the evolutionary characteristics of the original ALO. And repair operator, based on greedy strategy, is used to strengthen local search ability to revise infeasible solutions and optimize feasible ones. Finally, different dimensional 0-1 knapsack problems with the relationship between weight and profit of the items are carried out, and the comparison results with some existing algorithms demonstrate the effectiveness and stability of the proposed algorithm.

ABSTRACT. The rapid evolution of the automotive industry toward electrification and intelligence has generated massive volumes of news, creating a critical need for accurate multimodal classification. Existing methods, however, often fail to distinguish the unique semantic roles of headlines and bodies, leading to suboptimal cross-modal alignment. To address this, we propose the Multi-Scale Semantic Equilibrium Learning Transformer (HBI-Trans). This framework employs a Bidirectional Co-Attention (BCA) mechanism to capture hierarchical interactions between headline–image and body–image pairs, ensuring both coarse-grained and fine-grained alignment. Additionally, a novel Semantic Equilibrium Learning (SEL) mechanism utilizes dynamic calibration to balance headline and body contributions, effectively mitigating model bias. We further introduce AutoNews-MM, the first large-scale multimodal dataset in this domain, containing 20,081 samples. Experimental results demonstrate that HBI-Trans outperforms mainstream baselines with a 7.8\% improvement in macro-averaged F1 score, establishing a new benchmark for intelligent automotive information management.

ABSTRACT. This work addresses a challenging problem of estimating the 3D hand pose from a single RGB image. Most current methods regress the 3D hand pose from 2D key points, which are obtained from 2D heatmaps regressed from 2D image feature. Due to the single scale of the image feature and the loss of much information during the 2D key point regression, the accuracy of hand pose estimation is affected. To this end, we propose a hand pose estimation method based on multi-scale feature fusion and 3D convolutional neural network. This method consists of two parts. The feature extraction part extracts multi-scale features, and then fuses them through upsampling to obtain 3D image feature. The pose estimation part performs 3D convolution on the 3D image features to regress 3D heatmaps, and then obtain the 3D hand pose from them directly. Experimental results show that the proposed method has a mean per joint position error of only 7.32 mm on the Stereo Hand Pose Tracking Benchmark, which is better than most 3D hand pose estimation methods.

ABSTRACT. A hybrid ocean current energy harvester integrating an electromagnetic generator (EMG) and a triboelectric nanogenerator (TENG) is presented. The EMG employs a gear-accelerated four-motor structure to enhance low-speed electromagnetic conversion, while the nylon–Polytetrafluoroethylene TENG provides supplementary output and stabilizes the hybrid system. At 120 rpm, the device delivers 518.4 mW (231.43 W/m³) and a regulated output of approximately 5 V. Water-tank and lake field tests demonstrate reliable charging of a lithium battery, stable powering of sensors, and consistent operation under naturally varying flow conditions. The results confirm that the proposed hybrid generator offers an effective and robust power solution for long-term ocean monitoring.

ABSTRACT. Sending live video underwater is difficult due to very low bit-rates, high latency, and burst errors. This greatly limits how well underwater robots can see. We present ULTRA-VID, a real-time video coding and streaming method designed for bandwidth-limited underwater links. Based on DCVC-RT, our method improves network design, training, and packet handling for 8 frames. We use training techniques that imitate challenging underwater conditions to make video predictions stable at low bitrates. Our adaptive entropy model reduces data complexity based on available bits, helping to avoid errors when transmissions are interrupted. A simple self-attention module combines features from different scales quickly, improving the video quality even when textures are weak or movements are large. Frames are encoded and sent through a single low-bandwidth link, allowing real-time decoding on the receiving end. Tests on public datasets show ULTRA-VID reduces BD-Rate by 34\% compared to VTM and by 7\% compared to DCVC-RT at similar quality. A real-world test streamed 8-frame video groups at $128\times64$ pixels over a 100\,m underwater link, proving ULTRA-VID works effectively for underwater robot applications.

ABSTRACT. Precise and generalized mapping is the foundation of mobile robot navigation. Traditional Simultaneous Localization and Mapping (SLAM) algorithms rely heavily on environmental features, leading to instability, drift, or tracking loss in scenarios with illumination changes, shadows, or weak textures. While the state-of-the-art 3D vision foundation model, Visual Geometry Grounded Transformer (VGGT), enables direct regression of 3D attributes, its application to large-scale mapping is hindered by excessive computational costs due to global token interaction and cumulative drift in long sequences. To address these bottlenecks, this paper proposes a fast map generation framework based on VGGT. We introduce a Grouped Mapping Strategy with a Token Merging mechanism to significantly reduce memory overhead and accelerate inference. Furthermore, we design a hierarchical optimization scheme integrating Local Tracking Bundle Adjustment (BA) and a Global Loop Closure detection mechanism based on Aggregated Selective Match Kernels (ASMK). Experiments on 7-Scenes, TUM-RGBD, and ETH3D datasets demonstrate that our framework achieves superior pose estimation accuracy and dense reconstruction completeness while enabling efficient inference on resource-constrained devices.

ABSTRACT. Real-time Unmanned Aerial Vehicles (UAVs) object detection remains a highly challenging task.Light-weight attention mechanisms have been widely adopted to capture spatial context information and correlation between channels, aiming to improve detection accuracy while maintaining high inference speed and detection efficiency, achieving real-time detection in complex UAV scenes.However, existing methods jointly model spatial context information in horizontal and vertical directions, and independently but symmetrically encode directional information. Such designs overlook the inherent anisotropy induced by oblique perspective imaging.To address this issue, we introduce Anisotropic Interactive Attention (AIA), a light-weight and plug-and-play attention module for real-time UAV object detection.Specifically, AIA preserves directional spatial information through coordinate information embedding, and introduces an anisotropic dimension reduction strategy to flexibly allocate the representation capacity in the horizontal and vertical directions. Furthermore, AIA incorporates a cross-direction interaction mechanism to explicitly model the complementary correlations between the two directional branches before attention generation.Extensive experiments on VisDrone and UAVDT benchmarks demonstrate that AIA consistently improves the detection accuracy with only marginal additional parameters and computation cost, showing its effectiveness and practicality for real-time UAV object detection.

ABSTRACT. Multi-step facial manipulation often leaves weak local artifacts across spatially separated regions and introduces complex dependencies among editing operations. Although existing image-to-sequence detectors can recover potential manipulation categories and their execution order, standard Transformer encoders may attenuate fine-grained boundary and texture anomalies during global feature interaction. Moreover, a single-center spatial prior has limited flexibility in representing multiple discontinuous manipulation regions. To address these issues, we propose DAT-SeqFakeFormer, a dual-attention sequential deepfake detector with multi-scale information. In the encoder, Local Trace-Enhanced Self-Attention (LEA-SA) combines global multi-head self-attention with a parallel local convolutional path, helping preserve boundary discontinuities, texture perturbations, and high-frequency anomalies while maintaining long-range dependency modeling. In the decoder, Multi-Center Mixture-of-Gaussians Cross-Attention (MoG-CA) dynamically predicts multiple Gaussian centers and associated spatial scales from the current decoding state. The resulting masks are aggregated by position-wise maximum to form a unified multi-region spatial prior that guides cross-attention toward multiple candidate manipulation regions. Experiments on the \emph{Facial\_attributes} and \emph{Facial\_components} subsets of Seq-DeepFake show that DAT-SeqFakeFormer achieves the best results among the compared methods under both ResNet-34 and ResNet-50 backbones. With ResNet-50, it obtains 69.56\%/50.72\% Fixed-Acc/Adaptive-Acc on \emph{Facial\_attributes} and 73.58\%/56.86\% on \emph{Facial\_components}. Ablation studies further confirm the individual effectiveness and complementarity of LEA-SA and MoG-CA.

ABSTRACT. Small underwater objects are difficult to detect because visual degradation and limited pixel support weaken discriminative features and increase sensitivity to localization errors. This paper presents a controlled empirical study of two related aspects of this problem. The architectural study retains a P2 feature map. CARAFE and BiFPN are used for content-aware feature reassembly and bidirectional cross-scale feature fusion, respectively. LSK and the Efficient Multi-Scale Attention module are also examined under the same training protocol. The best-performing architecture increases average precision (AP) from 66.775 to 67.593 on the DUO benchmark. The resolution and localization study shows that increasing the input resolution from 640 to 768 pixels raises AP by 1.252 percentage points, from 66.775 to 68.027. Adding a normalized Wasserstein distance (NWD) term under the matched 768-pixel setting changes AP from 68.027 to 68.073, a marginal increase of 0.046 percentage points whose stability requires repeated runs. The two studies are not combined in a single model. Their results indicate that architectural design, input resolution, and the localization objective should be assessed individually. Among the settings tested here, input resolution produces the largest measured gain, whereas NWD has a limited additional effect.

ABSTRACT. Multiview point cloud registration recovers the absolute poses of all views by integrating pairwise transformations. While high-overlap pairs typically yield accurate transformations, low-overlap pairs are inherently less reliable. Existing methods often employ feature-space similarity to identify frame pairs for pose graph construction but may fail to assess a satisfactory overlap confidence. To address this limitation, we introduce graph-theoretic techniques that evaluate spatial compatibility (SC) within correspondence sets. Specifically, using biconnected components (BCC) to identify consistent consensus, we assign each frame pair an overlap confidence prior, enabling reliable pose graph construction. To improve efficiency, we employ Generalized-Mean Pooling for coarse retrieval before overlap estimation. For pose synchronization, we enhance Iteratively Reweighted Least Squares (IRLS) with a robust reweighting strategy incorporating accumulated residuals and an early stopping mechanism. Our approach can reject outliers efficiently through powerful overlap confidence prior and robust global pose optimization. Compared to state-of-the-art approaches, our method improves registration recall by 1.4\% on 3DLoMatch and achieves relative reductions of approximately 23.5\% in mean rotation error and 28.5\% in median rotation error on ScanNet.

ABSTRACT. Polarization cameras provide observations that are unavail able to conventional intensity imaging and can therefore improve the separation of target radiance from medium-induced path radiance. The separation is not intrinsic to polarization alone: it depends on the illumi nation state, target depolarization, water properties, analyzer sampling, and the reconstruction of the raw polarization mosaic. This review ex amines underwater polarization restoration as a coupled computational camera problem. A structured analysis of 45 primary and supporting studies is organized by the mechanism that makes the inverse problem tractable, the assumptions imposed on target and backscatter polar ization, and the solver used to recover radiance. The analysis covers polarization-difference and optical-rejection methods, global and spa tially adaptive physical inversion, circular and Mueller configurations, supervised and untrained networks, and physics- and domain-constrained learning. It also relates division by time, amplitude, aperture, and focal plane to errors in Stokes parameters and, consequently, to restoration bias. Across the literature, apparent progress is frequently confounded by unmatched water conditions, cameras, reference construction, and pro cessing latency. We therefore formulate an evaluation chain that requires acquisition disclosure, radiometric and polarimetric fidelity, system-level cost, and downstream task evidence. The review identifies three priorities for field deployment: uncertainty propagation from raw measurements, identifiable treatment of complex target polarization, and cross-water and cross-camera validation with reproducible ground truth.

ABSTRACT. Existing video super-resolution methods still suffer from fixed temporal propagation strategies and high computational costs of opti cal flow estimation modules. To address the issues of fixed propagation strategies and high computational costs of optical flow estimation mod ules in existing video super-resolution methods, this paper proposes a Motion-Aware Adaptive Propagation Network (MAPNet). Specifically, a Lightweight Efficient Optical Flow (LEOF) module and a Motion Com plexity Perception Module (MCPM) are designed. The proposed method reduces model complexity through lightweight optical flow estimation and generates adaptive propagation weights based on motion complexity information, thereby enabling dynamic fusion of temporal features. Fur thermore, the proposed motion-aware propagation mechanism enhances detail restoration in regions with complex motion, suppresses error prop agation, and improves the utilization efficiency of spatio-temporal infor mation. Experimental results on mainstream datasets demonstrate that the proposed method achieves superior video reconstruction performance and temporal consistency while maintaining low parameter consumption and computational cost.

ABSTRACT. To enable convenient and contactless control of smart furniture, this study de-velops an SSVEP-based brain–computer interface using the EMOTIV EPOC X headset. The system takes intelligent sofa lifting and descending as the control object, constructs a dual-frequency visual stimulation paradigm, collects and preprocesses real-time EEG signals through EpocX equipment, realizes SSVEP feature recognition based on canonical correlation analysis (CCA), and con-verts the recognition results into control instructions to drive the sofa actuator. The experimental results show that the overall mean recognition accuracy of the proposed system for sofa lifting and descending instructions reaches 85%, and the average single instruction response delay is less than 2.1s. The system has the advantages of simple deployment, low cost and high real-time perfor-mance, which can effectively realize non-contact intelligent control of home furniture, and provides a feasible technical scheme for the popularization of lightweight household BCI equipment.

ABSTRACT. Replacing feed receivers in a large radio telescope requires a robot to align and dock a component whose mass and scale are far beyond those of conventional robotic insertion tasks. This paper presents a force-bounded vision-force coupled docking method for a receiver lower platform weighing approximately 2.5 t and measuring about 3.2 m in diameter. A reconfigurable hybrid docking platform combines orthogonal translation, yaw rotation, and four independently actuated vertical supports. Before contact, an eye-in-hand camera localizes distributed bolt features; a plane-and-circle model estimates the docking center and attitude, which are mapped to the platform axes. After contact is detected, the controller switches from position-dominant motion to distributed force balancing. High-force supports retract or pause while low-force supports advance, subject to a point-force safety threshold and a force-spread termination criterion. A cross-coupled servo layer suppresses mismatch among the four vertical axes. Full-scale evaluation showed that the platform sustained a 2.51 t load for one hour and subsequently executed pose adjustment, limited maximum docking force to 386.7 N against a 500 N requirement, and supported a downstream bolt-hole alignment error of 0.4 mm. The complete integrated mock disassembly/reassembly sequence took 4 h 33 min 19 s, followed by an on-site demonstration at FAST. The results establish the feasibility of staged multimodal docking for heavy equipment maintenance in confined environments.

ABSTRACT. Six-degree-of-freedom (6D) object pose estimation supplies the rigid transformation connecting visual perception to robot action. This review examines pose as an interface for robotic manipulation, organizing representative work by test-time prior, sensing, representation, and temporal support. It traces pose through calibration and grasp/task frames, then analyzes picking, assembly, transparent-object handling, and closed-loop operation. Explicit pose is compared with task-relative transforms, completed geometry, and direct grasp prediction. A persistent object frame remains valuable for reusable grasp libraries, CAD collision reasoning, verification, and multi-step world models; direct or task-relative state can remove unnecessary interfaces when only the next action or inter-object relation is required. We therefore argue for paired evaluation: protocol-appropriate geometric metrics should be reported with task completion, contact and collision outcomes, calibration, latency, uncertainty, and recovery.

ABSTRACT. The rapid expansion of Internet of Things (IoT) systems in smart cities faces persistent power supply challenges, particularly in environments where conventional energy sources like solar and wind are unavailable. This study presents a novel Multi-Degree-of-Freedom Hybrid Wave Energy Harvester (H-WEH) that integrates electromagnetic (EMG) and triboelectric (TENG) mechanisms to enable self-powered IoT operation in the monitoring of urban water areas and underground sewer systems. The system features a gravity pendulum-based EMG module for rotational energy capture and a spring-assisted TENG module for vertical motion conversion, achieving synergistic energy harvesting across three degrees of freedom. The H-WEH’s power density of 0.5 mW/cm³ exceeds that of most reported wave energy harvesters. Experimental results show peak outputs of 20.8 V (EMG) and 412 V (TENG) under 20 cm wave excitation, with an integrated power management system enabling simultaneous battery charging (0.125 W average) and sensor operation. Field tests verify the system’s capability to power LoRa-based wireless communication over 200 meters, demonstrating its potential for continuous environmental monitoring in smart city applications. This work provides a viable solution for sustainable power generation in wave-rich environments while addressing the critical challenge of uninterrupted energy supply for distributed IoT networks.

ABSTRACT. Weather time series contain periodic fluctuations, slow trends, and channel-dependent scales. This paper proposes TCDMixer, a lightweight time-frequency mixing model for Weather forecasting. TCDMixer applies reversible normalization, channel-independent reshaping, lightweight trend-seasonal decomposition, an rFFT-based seasonal encoder, and a low-rank temporal trend decoder. The two stream representations are combined by gated lightweight fusion, and the normalized forecast is formed by a segment decoder together with a coarse trend prior and direct stream predictions. On the Weather benchmark, TCDMixer achieves an average MSE of 0.211 across four horizons while using 0.490M to 0.972M parameters. Under our local reproduction protocol, it obtains lower average MSE than DLinear, PatchMixer, PatchTST, and Informer on the same dataset. Ablation results indicate that decomposition, the trend branch, and prediction-level add-ons are the main sources of improvement.

ABSTRACT. As the scale of video data continues to grow, traditional video coding methods that rely on manually designed modules are increasingly limited in modeling complex scenes, removing spatiotemporal redundancy, and preserving perceptual quality at ultra-low bitrates. Neural video compression (NVC) integrates feature transforms, quantization, entropy models, motion estimation, and reconstruction into a unified rate-distortion optimization framework, offering a data-driven paradigm for video compression. This paper surveys the development of NVC. We first introduce the basic components and optimization objective of end-to-end learning frameworks, with emphasis on residual coding represented by Deep Video Compression (DVC) and conditional coding represented by the Deep Contextual Video Compression (DCVC) series. We compare these routes in terms of motion modeling, context utilization, and latent probability modeling. We then discuss the development of generative video compression, reviewing how methods based on generative adversarial networks, vector-quantized variational autoencoders, and diffusion models use generative priors, discrete latent representations, and conditional generation to improve perceptual quality at low bitrates. Finally, we summarize current challenges in computational complexity, temporal consistency, cross-scene generalization, generation authenticity, and evaluation criteria, and discuss future directions for video communication, immersive media, and virtual reality.

ABSTRACT. With the integration of distributed generation into the distribution network, traditional reclosing devices face challenges, especially when failure to disconnect the grid after a fault leads to reclosing failure. To address this issue, a fault discrimination method based on a Bayesian-optimized random forest model and a fixed number of capacitor switchings is proposed. This method determines the fault type by switching in the additional capacitor and analyzing its voltage response differences. Additionally, the parameter search space of the Bayesian optimization is dynamically adjusted to optimize the random forest model, improving training efficiency and classification accuracy. Simulation results show that the method effectively reduces unnecessary capacitor switching and significantly enhances the system's classification accuracy and robustness, demonstrating high practical application value.

ABSTRACT. Underwater video super-resolution (VSR) plays a crucial role in marine exploration and inspection. However, real-world underwater deployments are often affected by a coupled double-degradation process. On the one hand, physics-driven underwater imaging degradations such as scattering, absorption, and turbidity significantly deteriorate visual quality. On the other hand, bandwidth-limited transmission introduces strong codec compression artifacts, including blocking and ringing. The coexistence of these two types of degradations fundamentally challenges the assumptions underlying mainstream VSR pipelines.

In compressed underwater scenarios, high-frequency patterns introduced by compression are easily misinterpreted as genuine textures. This often leads to progressive artifact accumulation during recurrent propagation. Meanwhile, motion estimation becomes unreliable on blurred and block-corrupted frames, resulting in alignment errors and severe ghosting artifacts in reconstructed videos.

To address these challenges, we propose F-DCRNN, a Frequency-Decoupled Controllable Recurrent Network specifically designed for underwater compressed video super-resolution. The core idea is to explicitly separate degradation sensing from content modeling and to adaptively control the restoration strength according to the inferred compression condition.

First, a Dual-Domain Perception module extracts spatial content features while simultaneously estimating a compact compression fingerprint from frequency-domain cues. This fingerprint serves as a global control signal that characterizes the compression severity. Second, we introduce Structure-Isolated Temporal Alignment, which enhances motion compensation robustness by incorporating lightweight structural guidance while preserving domain-consistent RGB-based flow estimation. Third, we design a Frequency-Valve Recurrent Block (FVRB), an optimization-inspired recurrent unit that performs controllable feature denoising through dynamic soft-thresholding. The threshold is jointly modulated by the compression fingerprint and spatial structure cues, enabling effective suppression of codec artifacts without sacrificing delicate underwater details.

Extensive experiments on underwater compressed benchmarks demonstrate that F-DCRNN consistently improves both reconstruction fidelity and perceptual quality over strong VSR baselines across diverse compression levels.

ABSTRACT. Underwater object detection is essential in marine surveying, fisheries assessment, and autonomous underwater robot navigation. How- ever, traditional object detection models struggle with forward-looking sonar images due to noise, blurred boundaries, low contrast, and high computational complexity, resulting in reduced accuracy and limited real-time performance.This thesis proposes CASA-YOLO, an improved underwater object detection model based on YOLOv8, optimized in fea- ture extraction, fusion, and detection head design. The original C2f module is replaced by a novel C2f-CAFormer-CGLU structure, where CAFormer integrates self-attention with convolution to enhance feature modeling, and CGLU utilizes channel gating to improve feature distinc- tion. SimAM attention mechanisms are applied to both Backbone and Neck structures, enhancing sensitivity to underwater targets and sup- pressing background interference. Additionally, an Auxiliary Detection Head is introduced to capture and fuse multi-scale features, improv- ing detection of small and low-contrast targets.Experiments on a public sonar dataset show CASA-YOLO achieves a mean Average Precision (mAP@0.5:0.95) of 49.4%, a 3.6% improvement over YOLOv8, while maintaining computational efficiency. CASA-YOLO thus offers an effec- tive solution for underwater detection and monitoring, supporting more precise and efficient marine exploration.

ABSTRACT. Six-degree-of-freedom (6D) object pose estimation is essential for robotic manipulation and augmented reality. Existing methods often rely on mesh or point cloud representations, which limit differentiable rendering and generalization. We propose to integrate differentiable Gaussian Splatting into the FoundationPose framework to improve 6D pose estimation and reffnement. Objects are represented by neural Gaussians; we introduce front/back surface prediction to explicitly separate visible and self-occluded surfaces, improving geometric consistency and pose stability under occlusion. We then establish 2D–3D correspondences between image features and 3D model points, providing reliable constraints for PnP and iterative reffnement. Differentiable Gaussian rendering yields RGB, depth, and masks consistent with the observations, on which depth-based pose registration is performed; when depth is unreliable or registration fails, Gaussian rendering is used to synthesize depth and masks for robustness. A reffnement network predicts rotation and translation residuals from RGB and geometric features and is trained end-to-end with cropped data from Gaussian rendering. Experiments on Sam6D and MegaPose-style datasets show that the method preserves the generalization of FoundationPose while improving pose accuracy and robustness through front/back surface prediction and 2D–3D correspondence.

ABSTRACT. With the widespread use of 3D point cloud data in scene understanding, effi-ciently extracting features from point clouds and performing instance segmen-tation has become a key research focus. Due to the sparsity and disorder of 3D point clouds, traditional 2D convolution operations face many challenges when applied to point cloud processing, especially with voxelization, which can cause loss of point information and local texture features. To address these is-sues, this paper proposes a 3D point cloud instance segmentation algorithm based on sparse convolution and proposal generation, aiming to improve the accuracy of point cloud feature extraction and instance segmentation. The al-gorithm consists of two stages: In the first stage, sparse convolution is used to design a sparse convolution-based U-Net backbone network to extract deep features from the point cloud and generate instance proposals. In the second stage, the generated instance proposals are refined. A smaller U-Net network is used to further extract features and predict the category, instance mask, and mask score for each instance. By correcting classification errors and suppress-ing background, the model's accuracy and robustness are improved. Experi-mental results show that this method effectively enhances instance segmenta-tion performance in complex scenes, with promising application prospects.

ABSTRACT. Deep-sea Cobalt-rich crusts (CRC) serve as significant potential sources of critical strategic metals for high-performance energy storage batteries, for which Underwater Laser-Induced Breakdown Spectroscopy (LIBS) is recognized as an ideal in-situ detection technology. However, mining-induced turbidity fluctuations cause severe attenuation and distortion of spectral signals, rendering traditional quantitative methods inadequate. To address this challenge, this study proposes a novel Multi-Scale Dilated Adaptive Selection Kernel Network (MS-DASK-Net). By integrating convolutional branches with different dilation rates to construct dynamic receptive fields and utilizing attention mechanisms to adaptively adjust feature weights, the model achieves precise feature extraction of trace elements. Experiments demonstrate that compared with traditional Partial Least Squares (PLS) models, MS-DASK-Net exhibits superior robustness against interference under varying turbidity conditions, effectively overcoming the issue of non-linear spectral attenuation. This study validates the immense potential of deep learning in complex underwater in-situ analysis, providing key technical support for intelligent deep-sea exploration equipment.

ABSTRACT. The intelligent development of underwater operational equipment is a critical technological pillar for advancing the sustainable exploitation of marine resources and deep-sea scientific research. Soft gripping hands, leveraging the inherent compliance and adaptive deformation capabilities of their constituent materials, demonstrate unique advantages in unstructured underwater environments. This paper systematically reviews the technological evolution from terrestrial soft grippers to underwater soft gripping hands. It first recounts the early develop-ment and technical foundations of soft grippers, covering core topics such as ac-tuation mechanisms and variable stiffness methods. It then provides a detailed ac-count of domestic and international research progress on anthropomorphic soft dexterous hands, analyzing the critical role of rigid-soft coupling structures in en-hancing soft hand performance. Building on this foundation, the paper focuses on reviewing the structural design, system integration, and latest achievements of underwater soft gripping hands, including anthropomorphic underwater dexter-ous hands. It further discusses multimodal perception technologies and intelligent algorithms tailored for underwater scenarios. Finally, it summarizes typical appli-cations of underwater soft gripping hands in areas such as biological sampling, archaeological salvage, debris cleanup, and aquaculture harvesting, and offers an outlook on future development directions in this field.

ABSTRACT. Zero-shot 6D object pose estimation aims to acquire the translation and rotation of target objects in RGB-D images without the need for object-specific retraining. However, existing methods often suffer from severe performance degradation in cluttered scenes due to heavy occlusion and the lack of robust mechanisms for high-precision pose refinement. To tackle these challenges, we propose GRPose, a novel coarse-to-fine framework. GRPose introduces a geometric scoring-based visibility-aware segmentation module to reliably isolate target objects even under severe occlusion. Subsequently, to achieve highly accurate alignment, we design a cascaded three-layer recurrent pose refinement module powered by Transformers. Furthermore, a parameter smoothing strategy is integrated into the recurrent updates to effectively stabilize the residual predictions and prevent oscillation. In seven core BOP challenge datasets, GRPose achieves a state-of-the-art mean Average Recall AR of 70.5%. Extensive experiments demonstrate the superiority of our approach.

ABSTRACT. Global ocean plastic pollution has surpassed 150 million tons and increases by 8–12 million tons annually. Because traditional manual and trawl-based cleanup approaches lack sufficient single-machine salvage capacity, consume high energy, and have limited operational range and operating radius, they cannot meet expanding cleanup demands. To address this, we present an innovative autonomous underwater trash-cleaning robotic system comprising four mechanical subsystems—buoyancy, collection, clamping, and propulsion—and a double closed-loop PID controller featuring a gyroscope-based positional outer loop and an incremental velocity inner loop. Our visual recognition module builds on lightweight YOLOv5n, integrating GhostNetV2, BiFPN, and RFCBAMConv, trimming parameters from 2.50 M to 1.67 M and raising precision from 0.948 to 0.959 while maintaining $\mathrm{mAP}_{50}$ and $\mathrm{mAP}_{50\text{-}95}$. Ablation, real-world deployment, and comparative experiments demonstrate that our improved detector surpasses mainstream architectures such as DETR, EfficientDet, and CenterNet in detection performance. The complete robotic design provides a practical, scalable, energy-efficient, and environmentally friendly solution for large-scale marine plastic debris removal.

ABSTRACT. Underwater robots operating in complex marine environments rely heavily on visual perception; however, decoded video frames at the receiving end often suffer from severe degradation, including light attenuation, color shifts, and detail blurring. These issues are driven by the inherent physical properties of water absorption and scattering, and are further exacerbated by the extreme bandwidth constraints of underwater acoustic links. To restore perceptual quality post-transmission and empower downstream visual tasks, this paper proposes an innovative Multi-Condition Underwater Image Enhancement network based on Rectified Flow (RF-MCUIE). First, an Adaptive Masking (ADMK) strategy is proposed to dynamically isolate artifact-heavy training regions via color-space discrepancy measurement, thereby mitigating error propagation and significantly bolstering the network's generalization capability. Second, physical priors—namely, background light and transmission maps—are explicitly embedded as conditional guidance, constraining the color correction and dehazing processes to strictly adhere to real underwater optical mechanisms. Finally, to overcome the prohibitive inference latency inherent in traditional diffusion models, we formulate the generative process within a Rectified Flow framework. By learning a straight-line probability transport path from a noise distribution to the target clear image manifold, RF-MCUIE achieves high-fidelity generation in remarkably few inference steps (N=3). Extensive evaluations on the LSUI, UIEB, and U45 datasets demonstrate that RF-MCUIE establishes a new state-of-the-art across comprehensive quantitative metrics (e.g., PSNR, SSIM, LPIPS, FID) and subjective visual quality. The proposed method effectively neutralizes color deviation and restores structural details, striking an optimal balance between restoration quality and computational efficiency for bandwidth-constrained underwater acoustic communication systems.

ABSTRACT. Addressing the challenges faced by the Dynamic Window Approach (DWA) in path planning for Maritime Autonomous Surface Ships (MASS) under high wave conditions, specifically the inability to maintain safe encounter distances and insufficient route recovery post-avoidance, this paper proposes an Enhanced Dynamic Window Approach (EDWA). Initially, a high-fidelity container ship model and a virtual ocean environment were constructed using UE5, providing a reliable validation platform for subsequent algorithm testing. Subsequently, the A* algorithm was applied to determine the global path and set key waypoints. By incorporating waypoint guidance and sea conditions, route recovery factors and wave disturbance coefficients were integrated into the DWA evaluation function, and the sampling space was optimized based on COLREGs. Finally, the proposed EDWA algorithm was validated through the simulation platform. Under sea state level 3, the vessel successfully executed collision avoidance maneuvers in compliance with COLREGs while maintaining a safe distance, and swiftly and accurately re-stored the predetermined route post-avoidance. This demonstrates the algo-rithm’s superior robustness and reliability, underscoring its significant im-portance in ensuring the navigation safety of MASS.

ABSTRACT. With the advancement of embodied intelligence, the autonomous interaction capability of humanoid robots in unstructured environments has become increasingly critical. Stable, real-time visual object tracking is a fundamental prerequisite for achieving complex interactions. However, deploying high-precision tracking algorithms on edge devices faces significant challenges due to severe body oscillations during dynamic walking and the strict constraints on onboard computational power and energy consumption. To address these issues, this paper proposes and implements a vision-based single object tracking system for humanoid robots based on a heterogeneous dual-mainboard architecture. The system achieves hardware-software decoupling by utilizing an Intel Express for motion control and an NVIDIA Jetson for perception. Computationally intensive SOT algorithms are offloaded to the perception edge node, performing inference on RGB streams from a RealSense depth camera. Furthermore, a low-latency remote control interface was developed to support real-time video streaming and dynamic Region of Interest selection, establishing a complete "selection-tracking-servo feedback" closed loop. Experimental results on an actual humanoid platform demonstrate that the architecture effectively resists viewpoint perturbations caused by robot gait. While maintaining a high tracking success rate, the end-to-end latency meets the real-time requirements for high-frequency servo control, demonstrating significant engineering value.

ABSTRACT. Syndrome classification in Traditional Chinese Medicine (TCM) is a challenging multi-label prediction task, as patients often present with multiple coexisting syndromes and diverse clinical manifestations. Although deep learning and graph-based models have been applied to predict TCM syndromes, they still struggle to capture the diverse dependencies between nodes and the differences in association strengths. To overcome this limitation, we propose CGA-MLF, a Cross Graph Attention-driven Multi-Label Learning Framework for TCM syndrome prediction. The proposed framework constructs a multi-granularity multi-entity edge-weighted graph from Electronic Medical Records (EMRs), in which symptoms, disease locations, pathological attributes, syndrome nodes, and primary disease are jointly represented. Weighted edges are constructed according to symptom similarity, symptom–location matching, symptom–attribute associations, and symptom–syndrome relevance. Then, a multi-layer graph attention network is employed to learn the node representations from heterogeneous neighborhoods, while edge-weighted message passing is used to incorporate the strengths of clinical associations, and residual connections are introduced to preserve node representations across graph attention layers. Moreover, syndrome-node mask pooling is used to generate syndrome-oriented graph-level representations for multi-label prediction. Experiments conducted on the real clinical data set of TCM have shown that the performance of CGA-MLF is superior to the existing multi-label syndrome prediction methods. These results indicate that explicitly modeling multi-granularity entities and focusing on syndrome-oriented representations are useful for improving multi-label TCM syndrome prediction.

ABSTRACT. With the rapid development of Deepfake generation techniques, especially the emergence of new generative models such as diffusion models, forged videos have achieved remarkable improvements in texture detail, semantic consistency, and visual naturalness. As a result, obvious artifacts in high-quality Deepfake videos are becoming increasingly subtle, posing greater challenges to detection methods. To address the insufficient sensitivity of traditional methods to subtle anomalies in high-quality forgery scenarios and their limited ability to model complex temporal logic, this paper proposes DMT-Net, a Deepfake video detection network based on physical prior guidance and dual-scale temporal modeling. The proposed method constructs physical prior masks to guide the model toward potential tampered regions, extracts high-resolution shallow features to preserve fine-grained texture cues, and models interframe inconsistencies at both micro and macro temporal scales. Experiments on the FaceForensics++ dataset show that DMT-Net outperforms the YOLOv11 baseline under Raw, HQ, and LQ quality settings. In particular, under the LQ setting, the average ACC and AUC are improved by 0.75 and 0.62 percentage points, respectively, demonstrating the robustness of the proposed method in compressed degradation scenarios.

ABSTRACT. The core challenge of federated semi-supervised lung tumor segmentation lies in accurately capturing small and irregular tumor regions with blurred bound aries. This task is further constrained by patient privacy protection and lim ited expert annotations across multiple medical centers. Existing federated semi supervised methods improve decentralized training, but their segmentation back bones may still suffer from spatial information loss and anatomical background interference. To address these issues, we propose a dual-attention guided federated semi-supervised learning framework for lung tumor segmentation. Our method redesigns the local client backbone by introducing Attention Gates and Coordi nate Attention into the segmentation network. Specifically, Attention Gates are integrated into skip connections to suppress irrelevant anatomical responses and enhance tumor-related feature propagation. Meanwhile, Coordinate Attention is embedded at the bottleneck to encode long-range positional information along hor izontal and vertical directions, thereby improving the localization of small lesions. Experiments on four independent clinical datasets demonstrate that our method achieves competitive overlap performance and substantially improves boundary aware metrics, including HD95 and ASSD, compared with existing federated semi supervised methods.

ABSTRACT. Federated semi-supervised learning can alleviate annotation scarcity in medical image segmentation by leveraging both labeled and unlabeled data from multiple clients while preservingdataprivacy. However,pseudo-labelnoiseandnon-independentandidentically distributed(non-IID)datacauseclientreliabilitytofluctuateacrosscommunicationrounds, making it difficult for existing adaptive aggregation methods to accurately estimate client importance.Toaddresstheaggregationinstabilitycausedbyfluctuationsinclientreliability infederatedsemi-supervisedmedicalimagesegmentation,thispaperproposesaTemporal Reliability-aware Dynamic Aggregation strategy (TRDA). TRDA introduces historical relia- bility modeling into client weight estimation, uses an exponential moving average to smooth current-round performance fluctuations, and dynamically adjusts aggregation weights by jointlyconsideringcurrentsegmentationperformance,historicalstability,anddatavolume. For unlabeled clients, TRDA assigns a controlled small aggregation weight after the warm-up stagetoreducethenegativeimpactofpseudo-labelnoiseontheglobalmodel. Inthisway, TRDA provides a more stable estimation of client contributions, alleviates weight oscillations caused by data heterogeneity and local training randomness, and thereby improves the ag- gregationrobustnessandsegmentationperformanceofthefederatedsemi-supervisedlung noduleCTsegmentationmodel. ExperimentsonfourlungnoduleCTclientsdemonstratethat FedTRDA achieves superior segmentation performance compared with multiple federated semi-supervised baseline methods.

ABSTRACT. With the rapid advancement of medical imaging technologies, computed tomography (CT) imaging has played an increasingly vital role in clinical diagnosis. However, CT images are highly susceptible to various artifacts during the acquisition and reconstruction processes, such as respiratory motion, metal, and streak artifacts. These artifacts can severely degrade image quality and compromise diagnostic accuracy for clinicians. To address this issue, this paper proposes an artifact classiffcation method for medical CT images based on an improved EfffcientNet framework. Built upon the baseline EfffcientNet architecture, the proposed method incorporates a tailored dual attention mechanism for medical CT imaging and a dual-pooling feature fusion strategy, along with targeted optimizations to data augmentation and training strategies. Experimental results demonstrate that the improved CT-EfffcientNet model achieves a classiffcation accuracy of 97.70% across seven categories of CT image artifacts, outperforming the baseline EfffcientNet-B0 model (93.76%) by 3.94 percentage points. Furthermore, compared with several state-of-the-art mainstream models, the proposed method exhibits superior performance in terms of accuracy, thereby providing effective technical support for CT image quality control.

ABSTRACT. Dataset condensation aims to synthesize a compact training set that preserves the training utility of the original dataset. Distribution matching methods, especially M3D, achieve strong performance by minimizing maximum mean discrepancy between real and synthetic feature distributions. However, global representation-level distribution matching does not explicitly supervise low-level local statistics, such as color, texture, and edge responses, that are encoded in shallow feature maps. In this paper, we propose a simple loss-level enhancement that complements M3D with shallow feature moment matching. The proposed term matches the channel-wise mean and standard deviation of the shallowest encoder block between real and synthetic samples. It requires no architectural modification, no auxiliary network, and no additional bilevel optimization. Experiments on CIFAR-10 and CIFAR-100 with IPC=10 show that the proposed term consistently improves the reproduced M3D baseline, achieving 63.53±0.21% on CIFAR-10 and a short-run gain of 0.2% on CIFAR-100. We further compare SFMM with representative dataset condensation methods and provide ablation analysis on the matched statistics and layer choices. The results suggest that SFMM is a lightweight and complementary regularizer for M3D-based dataset condensation.

ABSTRACT. Video super-resolution (VSR) reconstructs high-resolution frames from low-resolution videos and requires both temporal propagation and detail restoration. Existing lightweight recurrent methods often rely on a compact shared state to simultaneously support propagation, detail preservation, and reconstruction, which can limit capacity under a restricted model budget and weaken the recovery of fine spatial details. To address this issue, we propose DPVSR, a lightweight recurrent VSR network with a dual-path design that separates temporal propagation from local detail preservation. A lightweight bidirectional recurrent branch captures temporal dependencies and inter-frame information flow, while a shallow detail branch retains current-frame appearance cues. We further introduce a lightweight post-fusion reconstruction module with spatial gating, residual reconstruction, and channel attention to enhance feature refinement before output generation. Experiments on REDS4 under the BI×4 setting show that DPVSR achieves a favorable balance between reconstruction quality and computational efficiency, demonstrating the effectiveness of the proposed dual-path propagation and reconstruction strategy.

ABSTRACT. Dataset Distillation (DD) learns a compact synthetic training set whose utility approximates that of a much larger real dataset. Distribution-matching approaches are attractive because they avoid the expensive nested optimization used by many gradient- and trajectory-matching methods. Hyperbolic Dataset Distillation (HDD) further introduces Lorentz geometry and aligns class-level hyperbolic centroids, providing a geometry-aware alternative to Euclidean feature-mean matching. Nevertheless, a single centroid primarily describes global class-level consistency and does not explicitly constrain representative intra-class regions. We therefore present Core-Aware Hyperbolic Dataset Distillation (CB-HDD), an extension of HDD that preserves the original full-centroid objective while adding a warm-started residual between radius-guided real and synthetic core prototypes. The implementation also supports optional class-center margin and radial-spread residuals, although the final configuration uses only the core term. Experiments on Fashion-MNIST, SVHN, and CIFAR-10 show that CB-HDD remains close to the HDD baseline across datasets and images-per-class settings, while the component study indicates that the core-only formulation is more reliable than the tested aggressive geometric combinations. These results position CB-HDD as a controlled structural extension of hyperbolic distribution matching rather than a replacement for the original HDD objective.

ABSTRACT. Rapid in-situ quantitative analysis of cobalt-rich crusts is important for deep-sea mineral exploration and mining. Underwater laser-induced breakdown spectroscopy (LIBS) provides a promising technique for multi-element detection, but sediment-induced turbidity can attenuate spectral signals, increase background fluctuation, and reduce the signal-to-noise ratio of weak emission lines. These effects make the quantitative prediction of Fe, Mn, and especially Co more difficult under turbid underwater conditions. To address this problem, this paper proposes a multi-scale adaptive feature learning method for turbidity-robust underwater LIBS quantitative analysis of cobalt-rich crusts. Cobalt-rich crust pellets were measured under clear-water, 50 FTU, and 100 FTU conditions. The proposed method uses multi-scale dilated convolutional branches to extract spectral features with different receptive fields and adopts an adaptive selection mechanism to fuse scale-dependent information. Its performance was compared with partial least squares regression and a conventional one-dimensional convolutional neural network. Experimental results show that turbidity causes obvious degradation in model prediction accuracy, particularly for the weak Co signal. Under 100 FTU turbidity, the proposed method achieved RMSEP values of 0.847, 0.912, and 1.526 wt.\% for Fe, Mn, and Co, respectively, outperforming both baseline models. The results demonstrate that multi-scale adaptive feature learning can effectively improve the robustness of underwater LIBS quantitative analysis under turbidity interference.

ABSTRACT. Integration of electroencephalography(EEG) based brain- computer interfaces (BCIs) with smart home technologies offers a transformative pathway for enhancing daily living, particularly for older adults and individuals with motor impairments. Following PRISMA 2020 guidelines, this systematic review synthesises 21 empirical studies published between 2020 and 2026, examining BCI paradigms, system architectures, and application scenarios in smart home environments. Our analysis reveals that SSVEP is the most prevalent paradigm due to its training-free nature and high information transfer rates (up to 37.4 bits/min), while hybrid systems combining multiple signals (SSVEP+EMG, EEG+EOG+ Gyro) are emerging to enhance robustness, with some achieving 100\% control accuracy. Deep learning models, particularly Transformer-based architectures, significantly outperform traditional classifiers, and consumer-grade EEG devices paired with edge computing platforms (Raspberry Pi) are lowering deployment barriers. Despite promising results, challenges persist including signal variability, limited command bandwidth, user fatigue, and ethical concerns regarding brain data privacy. Future research should prioritise lightweight deep learning, cross-subject transfer learning, digital twin integration, and long-term real-world evaluations to translate laboratory prototypes into practical, accessible assistive technologies.

ABSTRACT. Centralized semi-autonomous soccer systems—overhead camera, off-board computation, wireless-controlled differential-drive robots—are well- established platforms for multi-robot coordination research. However, the conventional monolithic software architecture tightly couples camera calibration, color-based perception, strategy generation, and communication framing, making iterative development slow and heavily field-dependent. This paper presents a modular framework that pairs a ROS 2 Humble upper-computer stack with an NVIDIA Isaac Sim digital twin for FIRA 5v5 small-size soccer. The key design principle is interface parity: physical and simulated environments share identical ROS 2 topic contracts, enabling the perception– strategy–control pipeline to be developed and debugged in simulation before hardware deployment. The vision module adopts Tsai two-step calibration, HSI color segmentation, region growing, and marker-based pose decoding for online robot and ball state estimation. The strategy module implements a three-layer rule library covering formation selection, role assignment, and differential wheel-speed generation. A multi-threaded Qt interface supports calibration, threshold tuning, match configuration, and visualization. Qualitative functional validation on Ubuntu 22.04 demonstrates robot recognition in both physical and simulated environments, physical shooting execution, and command- driven motion in simulation.

ABSTRACT. Autonomous driving technology needs to understand the pose information of the surrounding objects in the environment of the vehicle, but the complex scene and occlusion between objects in the scene bring challenges to the automatic driving technology. In this paper, we proposed a lightweight convolutional neural network for pose estimation, our method greatly reduces the amount of network parameters while ensuring the network detection performance. Specifically, we improved the feature extraction part of the original network, we used ResNet (Residual Network) pre-trained by ImageNet to extract image features, then passed the feature vector to the branch network for further processing, in the branch network we improved the original convolutional layer with 7×7 convolution kernel, the convolution layer with three convolution kernels of 3×3 is used instead. The quantity of parameters of our improved network is 38.7% of the original network, we trained on a synthetic data, experiment results showed that the improved network can ensure good performance on test data and accurately predict the pose information of objects in the scene.

ABSTRACT. Diffusion models have become a prominent way to represent multimodal robot actions, trajectories, and end-effector poses. This review summarizes their use in manipulation from three perspectives: policy formulation, geometric structure, and deployment. Conditional action diffusion avoids the mode averaging of unimodal behavior cloning and produces coherent action chunks that can be executed in a receding horizon. Recent methods extend this idea with goal conditioning, 3D observations, SE(3) equivariance, hierarchy, consistency distillation, foundation-scale pre-training, and visual-tactile feedback. The resulting policies are flexible and empirically strong, but iterative inference, limited hard-constraint guarantees, dependence on demonstration coverage, and inconsistent evaluation remain important obstacles. We compare representative methods and identify research priorities in few-step generation, safety-constrained denoising, contact-aware control, heterogeneous multi-robot training, uncertainty calibration, and reproducible benchmarking

ABSTRACT. Robotic peg-in-hole assembly is sensitive to incomplete depth observations, near-symmetric workpiece pose ambiguity, cali- bration errors, and residual pose deviations. This paper presents a CAD-prior vision-guided assembly pipeline for a UR7e robot. SAM 2 provides a workpiece mask, and FoundationPose estimates the 6D pose of the complete workpiece. Rather than tting the hole axis from incom- plete depth measurements inside the hole, the hole center and axis are derived from a xed CAD transformation between the workpiece frame and the hole-axis frame. Candidates for the near-symmetric workpiece are rst ltered by an upward-axis constraint in the robot base frame and then rescored using mask overlap, depth consistency, and projected contour error. The resulting hole-axis pose is transformed through eye- in-hand calibration to generate pre-insertion and insertion targets. A threshold-based force correction rule is evaluated separately in Isaac Sim as a preliminary simulation study. Under the tested conditions, usable segmentation masks were obtained in 96.7100% of images. Simulation increased insertion success from 70% to 95% and reduced average peak contact force from 36.2 N to 20.1 N. Real-robot experiments without closed-loop force correction achieved 31 successful insertions in 40 trials. These results support the feasibility of the visual localization and pose-mapping pipeline for the tested workpiece, while force correction remains a simulation-based preliminary component.

ABSTRACT. 3D mesh models are widely used in digital twins, cultural heritage digitization, and cloud-based 3D content services. Existing schemes, however, still show limitations in layered controllability, time efficiency, and inter-layer dependency, while flexible tiered access is often required according to user privileges. This paper proposes a controllable layered Arnold-DNA encryption scheme. The scheme regularizes vertex coordinates in double precision and constructs mantissa sub-blocks; generalized Arnold scrambling and DNA XOR diffusion are then combined for encryption, while a ciphertext-driven chain key mechanism strengthens inter-layer dependency. The proposed scheme supports flexible layered decryption and can progressively recover visual information according to user privileges. Experimental results show controllable layered visual recovery and favorable performance in distortion, timing, and attack-sensitivity evaluations.

ABSTRACT. Drug-drug interaction (DDI) prediction is crucial for facilitat- ing safe combination therapies and reducing unexpected adverse drug reac- tions. However, existing methods typically employ fixed weighting across all modalities. Consequently, it is a challenge to achieve self-adaptive multi-view fusion and feature calibration. In this study, we proposed FC-SMFL, a Feature Calibration-based Self-adaptive Multi-view Fusion Learning Model for DDI prediction. The model uses Transformer and Graph Isomorphism Network (GIN) encoders to extract 1D sequence features and 2D molecular graph features of drugs. It adaptively gen- erates independent weights for the 1D and 2D views via the gating networks and combines these views using a bilinear interaction term, thereby achieving prior multi-view fusion. Furthermore, we propose a feature calibration strategy that calculates a coefficient to adjust the proportion of the original drug features and generated features, effectively mitigating the over-smoothing problem. Experiments on Deng’s dataset demonstrate that FC-SMFL outperforms representative baselines, with a 3.37% improvement in F1-Score and consistent gains across all evaluation metrics.

ABSTRACT. Ultra-low-bitrate underwater video transmission requires both perceptually faithful reconstruction and secure delivery over bandwidthconstrained, potentially untrusted channels. This paper presents Secure PGTRP-VC, a security-aware neural video compression framework that integrates perception-guided latent coding with authenticated bitstream protection. A frozen Wan-VAE first maps RGB video into a compact spatio-temporal latent space, which is causally compressed by a DCVCRT-derived codec. A Perception-Guided Temporal Routing Prior selectively refines the predicted Gaussian mean using decoder-reproducible temporal context, improving perceptual quality without transmitting extra side information. To introduce codec-level key dependence, Keyed Latent Protection applies a reversible signed permutation to mean-centered latent symbols and their matched entropy parameters within causally independent coding groups. The resulting neural payload is further protected by AES-GCM, while sequence indices and codec metadata are authenticated to detect tampering, replay, deletion, and reordering. Experimental results show that the mean-only PGTRP configuration reduces P-frame BPP from 0.0089 to 0.0067 and improves DISTS and LPIPS from 0.1330 and 0.1018 to 0.1010 and 0.0530, respectively, with a moderate PSNR reduction. The framework provides a unified design for perceptual ultra-low-bitrate underwater video compression and secure transmission.

ABSTRACT. Accurate egg counting supports laying performance monitoring and farm man agement, yet line counting remains difficult when eggs undergo stop and go motion, near line rebound, skipped frames, and edge device limits. General trackers assume smoother motion and rarely combine stable association, reliable count triggering, and lightweight deployment. We propose a lightweight framework for stop and go egg sort ing lines that integrates two task specific components: a pure momentum Distance Intersection over Union (PM-DIoU) tracker and a production aware state machine counter. With YOLOv8n as detector, PM-DIoU uses pure momentum prediction and Distance Intersection over Union (DIoU) association, while outputting the current de tector box. This reduces box inertia and fragmentation during stops, restarts, and skipped frames. The state machine models initial ignored, near line, counted, and terminated states to suppress repeated triggers from rebound, track breaks, or iden tity reinitialization. Evaluation combined annotated tracking, count ground truth, low frame rate tests, long video validation, and edge benchmarking. On three anno tated videos, PM-DIoU with the counter achieved 99.8% tracking accuracy and one egg counting error. In 1 h 53 min 26 s of long video validation, the total counting error was also one egg (0.01%). Low frame rate tests gave mean count errors of 0.1% and 0.2% at 12 and 6 frames per second. Open Visual Inference and Neural Network Op timization (OpenVINO) reached 22.77 frames per second on an Intel N100 industrial computer. These results demonstrate deployable edge vision for stop-and-go egg-line counting.

ABSTRACT. Multimodal data fusion is of great significance in intelligent diagnosis, while existing fusion methods—relying on static weighting or global augmentation—fail to capture patient-specific inter-modal dependencies and often lead to over-smoothing. To address these limitations, we propose HDAF-CAD, a hybrid data augmentation and fusion-driven cross-modal adaptive distillation framework for intelligent medical diagnosis. This framework presents a cross-modal adaptive distillation mechanism, which dynamically adjusts the smoothness of the soft labels to achieve efficient teacher-student transfer and effectively alleviates the modality imbalance problem. Then, a multi-head cross-attention fusion module is designed, using CT as the query to retrieve complementary imaging/clinical features for patient-specific fusion. Additionally, a partial mixed data augmentation strategy is investigated, interpolating the sub-batches to balance the generalization ability and avoid excessive smoothing. The experimental results on the real-world MMIST-ccRCC dataset show that HDAF-CAD outperforms the baselines. In terms of BAcc, AUC, F1 score, and sensitivity indicators, it reached 0.7309, 0.8129, 0.8456, and 0.6000, which increased by 20.52%, 25.97%, 0.97%, and 54.67% compared with the best baseline, LUNAI-fCT. Our framework provides a promising paradigm of reliable prognosis, constructing a synergy model of adaptive knowledge transfer, fine-grained interaction, and controlled augmentation.

ABSTRACT. To address the strong non-stationarity, high noise, and insufffcient modeling of cross-stock dependencies in China’s A-share market, which is affected by policy uncertainty, sectoral linkage, and market ffuctuations, this paper proposes a Market-Guided Shared Mamba with CNN-based Temporal Attention model, named MGSM-CTA. The model takes individual stock historical features and market features as joint inputs, and introduces macro-level market states into stock-level temporal modeling through a market-guided mechanism. In this way, the model can capture the historical evolution patterns of individual stocks while considering the inffuence of overall market ffuctuations on the prediction target. Structurally, a market-guided shared Mamba module is ffrst designed to efffciently capture long-sequence dependencies through the state-space model and enhance common feature representation among different stock samples via shared parameters. Then, a CNN autoencoder-based temporal attention mechanism is introduced to adaptively generate temporal importance weights through an encoderdecoder structure, thereby highlighting key temporal patterns and suppressing ffnancial noise. Finally, a differentiated tokenization strategy is constructed to enhance feature representation from intra-stock, interstock, and temporal dimensions. Experimental results on the CSI300 and CSI800 datasets show that MGSM-CTA achieves competitive predictive and ranking performance in terms of IC, ICIR, RIC, RICIR, AR, and IR. Ablation experiments further verify the effectiveness of market information guidance, CNN-based temporal attention, and differentiated tokenization strategies.

ABSTRACT. Photovoltaic power forecasting is inherently challenging because plant out-put is governed by complex interactions between stochastic weather condi-tions and time-varying system operating states, resulting in nonstationary and multi-scale temporal patterns that are difficult for conventional models to capture. To address these issues, this paper proposes an STL-MHDA-iTransformer-BiLSTM model. The framework first employs STL decomposi-tion to separate the original signal into trend and seasonal components, which are then processed in dedicated branches. The trend branch uses a BiLSTM network to encode temporal dependencies bidirectionally, effective-ly capturing long-range sequential information. The seasonal branch adopts an MHDA-iTransformer with multi-head differential attention to extract fi-ne-grained periodic features and inter-variable correlations. An adaptive weighting module (AMS) subsequently fuses the two branch outputs to pro-duce the final forecast. Extensive experiments on a real-world photovoltaic plant dataset demonstrate that the proposed model consistently outperforms representative baselines in convergence speed, trend tracking, and overall forecasting accuracy. Moreover, ablation studies systematically confirm the individual contributions of the STL decomposition, MHDA mechanism, BiLSTM module, and AMS fusion strategy, validating the rationality of each design choice.

ABSTRACT. Electricity load forecasting is a critical enabler for the safe and stable operation of power systems and for achieving dynamic supply–demand balance. However, the nonstationarity of load fluctuations and the intricate interplay between trends and periodic patterns pose formidable challenges to forecasting accuracy. To address these issues, this paper proposes a hybrid model termed Informer-DFT-Mamba. The model applies the Discrete Fourier Transform (DFT) to analyze the original load sequence in the frequency domain, employs low-pass and high-pass filters to extract low-frequency and high-frequency components respectively, and reconstructs the corresponding time-domain trend and seasonal component sequences via the Inverse DFT (IDFT). The trend component is modeled by Mamba for long-range dependency capture, and the seasonal component is modeled by Informer with its probabilistic sparse attention mechanism for global periodicity extraction. Experimental results on real-world electricity load datasets demonstrate that the proposed method outperforms existing mainstream models, exhibiting stronger generalization ability and higher forecasting accuracy.

ABSTRACT. We reveal that attention mechanisms in object detection exhibit strong level-dependency: directional attention benefits high-resolution layers but not semantic layers. Based on this finding, we propose Scale-Specialized Directional-Channel Attention (SSDCA), a pyramid-level-aware heterogeneous attention strategy that assigns Crack-Oriented Directional Attention (CODA) to the high-resolution P3 layer for elongated crack modeling, and Squeeze-and-Excitation (SE) channel attention to the semantically richer P4 layer for large-object recalibration. CODA employs four oriented depthwise convolutions (horizontal, vertical, two diagonals) to capture anisotropic crack structures, initialized with diagonal patterns as a strong directional inductive bias.Experiments on a tunnel defect dataset with 3,185 images show that SSDCA improves YOLOv8n by 7.0 percentage points on the validation set (0.7133 to 0.7831, 5-seed mean, sign test p = 0.031) and 2.8 pp on the test set, with only 0.01M additional parameters and 0.1 ms latency overhead. Ablation confirms that the scale-specialized assignment (CODA@P3+SE@P4) outperforms the reversed assignment (CODA@P4+SE@P3) by 4.5 pp, validating the pyramid-level-aware design.

ABSTRACT. Mechanical component detection is challenged by large scale variation, reflective surfaces, clutter, occlusion, and high inter-class similarity. Lightweight convolutional detectors preserve local detail but have limited capacity to model long-range part configurations, whereas direct state-space block replacement can disrupt pretrained features when domain data are limited. To address these problems, we propose DGSSA-YOLO, which contains a Dual-Granularity Selective State Adapter (DGSSA) and a Mechanical Scale-Quality (MSQ) loss. DGSSA retains the pretrained C2f mapping and learns a zero-start residual from a native-resolution adaptive geometry branch and a pooled four-direction selective-state branch. An input-conditioned channel gate fuses the two branches, and two adapters are deployed at matched P4 stages in the backbone and neck. MSQ normalizes scale-aware localization weights and quality-hardness classification weights so that small and uncertain samples are redistributed without changing the nominal loss magnitude. Experiments on Mechanical Parts Dataset show that DGSSA obtains 89.61% mAP@0.5 and 74.22% mAP@0.5:0.95 with 3.07 M parameters. Relative to a direct AKCMamba adapter, it improves recall by 1.85 points and mAP@0.5:0.95 by 0.16 points while reducing parameters and computation. MSQ further raises test precision from 88.75% to 90.81% at essentially unchanged mAP@0.5:0.95, yielding a precision-oriented operating point for false-alarm-sensitive inspection.

ABSTRACT. Large-scale medical imaging databases are essential for developing generalizable medical AI systems, but their construction is often hindered by heterogeneous sources, duplicated records, inconsistent terminology, and institution-specific reporting styles. Direct image-level organization is costly and typically requires expert annotation. To this end, we propose a report-driven framework that uses radiology text as a scalable and interpretable proxy for the preliminary organization of heterogeneous imaging data. Using the Japan-Medical Image Database, which contains computed tomography and magnetic resonance imaging reports from more than 10 academic medical centers in Japan, we developed a domain-aware preprocessing pipeline that reduces textual noise, normalizes report content, and filters low-information entries, yielding 195,623 valid reports from 3,026,213 raw reports. The cleaned reports were embedded and clustered to construct a structured report-level corpus, with representative cases and clinical labels automatically derived for interpretation. Compared with clustering before preprocessing, our framework improved the comprehensive clustering score from 0.5820 to 0.6586 and increased the optimal number of clusters from 11 to 25. At their respective optimal clustering solutions (K = 11 for raw reports and K = 25 for preprocessed reports), the clustering metrics showed substantial improvements after preprocessing. The preprocessed reports also achieved consistently better performance across comparable values of K, indicating that the improvements were attributable to the proposed preprocessing strategy rather than solely to the change in the optimal number of clusters. These findings demonstrate that domain-aware preprocessing enables radiology reports to reveal clinically coherent structures in large-scale imaging data, providing a practical foundation for scalable database curation and future image-level or multimodal analysis.

ABSTRACT. Multimodal visual object tracking integrates complementary information from heterogeneous sensors, such as RGB-thermal, RGB-depth and RGB-event streams, to improve tracking robustness in challenging environments where RGB-only observations may become unreliable. Existing multimodal trackers, however, are often designed for specific modality pairs and therefore rely on separate fusion structures or adaptation parameters for different sensor combinations. We present MMLTrack, a unified multimodal tracking framework based on mixture-of-experts parameter-efficient tuning. The proposed tracker introduces a Meta-Fusion module to map RGB and auxiliary-modal features into a shared representation space, thereby enhancing semantic alignment and collaborative representation. To balance unified modelling capability and parameter efficiency, a multimodal MoELoRA module is further developed by freezing the Vision Transformer backbone and inserting low-rank adaptation experts with a routing mechanism. The resulting framework preserves the representation capability of the pre-trained RGB tracker while enabling modality-aware parameter selection for RGB-T, RGB-D and RGB-E tracking. Experiments on several datasets covering diverse modalities show that MMLTrack achieves leading performance across heterogeneous multimodal tracking benchmarks, demonstrating the effectiveness of the proposed fusion and parameter-efficient adaptation design.

ABSTRACT. Multimodal data fusion is essential for artificial intelligence-assisted diagnosis and healthcare systems. Medical images and their reports contain mutually complementary visual and semantic clinical information. However, most existing approaches fail to adequately exploit cross-modal semantic associations, erode transferable visual knowledge in the model adaptation phase, and suffer from limited generalization ability toward unseen disease categories. We propose SM-GHMRL, a Semantic Matching-Gated Hierarchical Multimodal Representation Learning framework for zero-shot multi-label medical image recognition. SM-GHMRL integrates complementary Findings-Impression semantics through gated fusion, combines hierarchical visual features, and constructs transferable disease prototypes via dynamic clinical prompting for zero-shot multi-label prediction without additional fine-tuning. Experiments on CheXpert show improvements of 8.01%-13.20% in AUC, 7.69%-21.56% in Accuracy, and 6.90%-21.67% in F1-score over three baselines. The results demonstrate its effectiveness and potential for supporting efficient zero-shot disease recognition and AI-assisted diagnosis in real-world clinical settings.

ABSTRACT. Deep-sea tracked mining vehicles operate on soft and spatially varying seabed sediment, where hydrodynamic resistance, sediment interaction, and asymmetric track slip complicate trajectory tracking. This paper develops a curvature- and propulsion-loss-aware adaptive fuzzy control framework in which a continuous reference-speed scheduler combines local path curvature with an encoder--navigation propulsion-loss index. A look-ahead-point transformation is integrated with adaptive fuzzy surge and yaw control, while the uncertainty model separates state-dependent nonlinearities from bounded time-varying residuals. The stability analysis incorporates the filtered virtual-control derivative used in the sampled implementation and establishes uniform ultimate boundedness of the look-ahead tracking, velocity, and fuzzy-weight errors within a prescribed compact operating domain. The method is evaluated numerically on a smoothed keyhole path with spatially fixed soft-sediment regions, exogenous asymmetric slip, measurement noise, slowly varying plant coefficients, and low-frequency current disturbance. A matched non-fuzzy robust baseline and four fuzzy ablations are compared under identical conditions. The proposed configuration achieves position and heading RMSEs of 0.074 m and 0.909°, respectively, and reduces normalized control effort by 6.8% relative to the non-fuzzy baseline. Relative to curvature-only scheduling, the position RMSE is reduced by 5.6%, with a 9.8% increase in completion time.

ABSTRACT. Face images are sensitive biometric data widely used in social networks, intelligent surveillance, mobile devices, and identity authentication systems. Unlike ordinary visual images, face images contain stable identity-related characteristics, such as inter-ocular distance, facial contour, facial component layout, and skin texture. Unauthorized acquisition, dissemination, or recognition of face images may therefore lead to serious privacy leakage and identity misuse. Existing visual privacy protection methods, including mosaicking, blurring, occlusion, conventional encryption, visually meaningful encryption, and AI-based face replacement, usually satisfy only part of the privacy requirements. They often fail to simultaneously achieve visual naturalness, original semantic preservation, identity obfuscation, and authorized recoverability. To address these limitations, this paper proposes a recoverable face privacy protection method that integrates CodeFormer and StegFormer. Specifically, the original face image is first degraded by a mosaic operation to weaken local identity-related textures. CodeFormer is then employed to perform blind face restoration on the degraded image, producing a visually natural face image with improved structure and texture. Subsequently, StegFormer embeds the original face image into the CodeFormer-restored image to generate a publicly shareable stego image with privacy protection capability. Authorized users can further recover the original face image through the reveal network. Experimental results demonstrate that the proposed method achieves a favorable balance among visual naturalness, semantic usability, identity privacy protection, and high-fidelity authorized recovery.