Abstract:To address intensified load fluctuations, enlarged peak-valley differences and capacity stress caused by large-scale electric vehicle integration in residential scenarios, this paper studies the capacity-constrained optimization of aggregated orderly charging. A grouped multi-objective optimization model is established to coordinate charging cost, total load fluctuation and charging behavior deviation, while considering group-wise energy conservation, power limits, capacity constraints and scenario feasibility. Based on a decomposition-based multi-objective evolutionary framework and success-history based adaptive differential evolution, a structure-enhanced congestion-guided method is proposed. Energy-preserving pairwise perturbation, congestion-guided evaluation and local refinement are introduced to improve the search capability under complex constraints. Mild, medium and highly capacity-tight residential scenarios are constructed for validation. The results show that the proposed method achieves the best overall performance in all five formal experimental settings. In the highly capacity-tight scenario, compared with the main benchmark method, the solution-set approximation error is reduced by about 26.1%, and the solution-set coverage quality is significantly improved. The proposed method obtains an average rank of 1 and outperforms the benchmarks in all five scenarios, with an average runtime of 47.8 s. Ablation results indicate that energy-preserving pairwise perturbation is the key mechanism for improving solution quality, while congestion-guided evaluation and local refinement further enhance the performance under tight capacity constraints. The results verify the effectiveness of the proposed model and method in achieving safe charging-load transfer, load smoothing, and charging-behavior preservation under residential capacity constraints.

Abstract:Data center cooling systems account for a substantial proportion of overall energy consumption and are significantly affected by workload fluctuations and thermal inertia, resulting in multiple disturbances and strict physical constraints in temperature control. Conventional control methods often struggle to balance safety and energy efficiency, while existing reinforcement learning methods may generate physically infeasible actions during exploration, limiting their practical deployment. To address these issues, this paper proposes a physics-informed safe reinforcement learning method for temperature control based on an open-source simulation environment. The temperature control problem is formulated as a Markov decision process. Under the existing definitions of the state and action spaces, a dynamic safety boundary model coupling dew point temperature, system stability and air enthalpy is developed. A differentiable projection mechanism is further embedded into the policy network, enabling differentiable safety mapping of actions in the continuous action space. In year-round simulations based on EnergyPlus, the proposed method reduces the number of temperature violations to 10, increases the proportion of cold-aisle temperatures within the recommended range to 99.2%, and limits the violation severity to within 0.03℃. Meanwhile, it significantly reduces control fluctuations and improves system operating stability. While ensuring safety, the method further improves system energy efficiency. The results demonstrate that, without modifying the system structure, the proposed method enables reinforcement learning policies to effectively represent and consistently satisfy complex physical constraints, achieving a favorable balance among safety, stability, and energy efficiency. This study provides a feasible approach for the practical deployment of reinforcement learning in data center temperature control.

Abstract:In the scheduling of source network load storage in the distribution network, the dual uncertainty of meteorology and load has complex temporal dependence characteristics. If the martingale difference sequence correlation between the two is ignored, it can easily lead to conservative mismatch and interface instability risks in scheduling decisions, resulting in voltage exceeding limits. Therefore, this study proposes a robust optimization scheduling method for source network load storage considering both meteorological and load uncertainties. Firstly, a first-order autoregressive time series correlation model for meteorological forecast errors and a box interval model for load fluctuations are constructed, and a net load coupling mechanism considering dual uncertainties is established to form a joint uncertainty model that is closer to actual extreme scenarios; then, a multidimensional box shaped uncertain set with adjustable conservatism is used to characterize the fluctuation domain of net load, and the goal is to minimize the total cost of system power purchase, energy storage aging and interruptible load compensation. Taking into account constraints such as power balance, voltage safety, energy storage state of charge and interruptible load, a robust optimization scheduling model is constructed. The main problem and sub problems are iteratively solved through the column and constraint generation algorithm. Finally, design a real-time collaborative response framework for adaptive charging and discharging of energy storage and flexible adjustment of interruptible loads to address residual deviations in actual operation. The simulation results show that this method maintains the voltage of all nodes in the safe range of 0.94~1.06 per unit (p.u.) under the worst-case uncertainty scenario, and the safety margin remains positive, verifying the superiority of this method in improving the safety of the distribution network.

Abstract:With the deepening reliance of the socio-economy on electricity, the threat posed by extreme weather events to power system stability has become increasingly prominent. Consequently, conducting research on power system resilience assessment holds significant practical value. This study aims to construct a resilience assessment framework for power systems that accounts for the spatio-temporal characteristics of extreme weather. First, an extreme weather model incorporating spatio-temporal characteristics is developed, where the typhoon model is dynamically corrected based on historical typhoon data and real-time geographic coordinates (latitude and longitude). Second, considering factors such as power system topology, operational status and disaster response capabilities, three comprehensive indicators are proposed to evaluate the system′s resilience level across three phases: pre-disaster prediction, disaster-time endurance and post-disaster recovery. Furthermore, stochastic modeling is performed using 15 years of measured typhoon data. The IEEE RTS-79 bus system is utilized for simulation to derive resilience baseline values. From the perspective of engineering practice, three resilience enhancement strategies—increasing standby capacity, implementing load transfer and improving maintenance efficiency—are proposed and verified. This assessment framework not only reveals vulnerabilities in power systems during disaster response but also provides theoretical basis and practical guidance for enhancing power system resilience, which is of great significance for ensuring national energy security and maintaining socio-economic order.

Abstract:In low-temperature environments, the battery of the substation inspection robot exhibits open circuit voltage hysteresis loop distortion, resulting in severe low-temperature drift accumulation in the estimated state of charge and significant non-stationary deviations in the predicted motion energy consumption results. Therefore, this study proposes a range enhancement strategy that integrates low-temperature discharge modeling, motion energy consumption prediction and layered compensation. Firstly, a semi empirical electrochemical model is established, and an extended Kalman filter is used to estimate the state of charge of the battery online, suppressing the cumulative drift caused by low-temperature nonlinear polarization; then, combining real-time state of charge and path planning information, output the remaining endurance time prediction result in a closed form; finally, a dual layer compensation architecture is designed. The upper layer dynamically reorders inspection task points based on the endurance tension factor and adjusts the charging trigger threshold, and the lower layer designs an energy controller based on the port controlled Hamiltonian theory, optimizing joint torque distribution through injection damping and energy shaping, experiments have shown that in low-temperature environments ranging from -18 to -8℃, this strategy can complete all 8 inspection points on a single machine, with a utilization rate of 76.2% for the state of charge, which is 43.1% to 70.5% higher than existing methods. The operation time is extended by 64.8%, and the instantaneous power variation coefficient is as low as 0.11. The current sample entropy is reduced to 0.52, verifying that this measurement can achieve accurate prediction and efficient compensation of endurance under harsh low-temperature conditions.

Abstract:For target tracking under conditions of inaccurate noise covariance and non-uniform sampling, this paper proposes an adaptive filter. During the measurement modeling phase, this filter incorporates an dimensionality expansion strategy to reconstruct the measurement model under non-uniform sampling, thereby effectively preventing measurement information omission and temporal mismatch issues. To achieve precise estimation of dynamic targets, this paper models the target state as an inverse Wishart distribution, where the target′s kinematic state is modeled as a Gaussian distribution and the noise covariance as an inverse Wishart distribution. By introducing a variational inference mechanism, a Bayesian recursive process is derived for optimal estimation of the target state under conditions of inaccurate noise covariance and non-uniform sampling. Simulation results validate the proposed algorithm′s robustness and estimation accuracy in complex observation environments. Compared to existing Kalman filters, it demonstrates significant advantages in handling inaccurate noise covariance and non-uniform sampling, exhibiting superior adaptability and tracking performance.

Abstract:To address the limitation in measurement accuracy caused by phase mismatch in traditional single-reference signal demodulation methods, a dual-channel synchronous demodulation method is proposed to enhance the precision of a precision printed-circuit board (PCB）-based electromagnetic induction-type angular displacement sensor. By providing independent square-wave reference signals for the sine and cosine output signals of the sensor, this method achieves attenuation-free demodulation through phase matching, fundamentally eliminating the gain inconsistency induced by phase mismatch. Experimental results show that, compared with the conventional method, the proposed approach reduces the original full-cycle measurement error from ±0.0323 5° to ±0.007 6°, lowers the compensated full-cycle error from ±0.006° to ±0.003 4°, improves the dynamic range of the sensor output signal from approximately ［-16 001, 9 338］ to about ［-30 000, 30 000］, and enhances angular stability to 0.000 1°. This study provides an effective technical solution for improving the performance of high-precision PCB-based electromagnetic induction-type angular displacement sensors, offering broad prospects for engineering applications.

Abstract:Addressing the issues of significant scale differences in defect targets, strong background interference and potential missed detection of small and occluded targets in aerial images of transmission lines, a poly feature fusion detector based on YOLOv11n ( PFD-YOLO) is proposed. Firstly, by incorporating poly kernel Inception block （PKI Block） from poly kernel Inception network （PKINet） to enhance C3K2, the model complexity is effectively reduced while the feature extraction capability is enhanced. Secondly, a feature focusing diffusion pyramid network (FFDPN) is designed to reconstruct the feature fusion structure on the Neck side. The feature focusing module (FF) facilitates efficient interaction and contextual information propagation among features of different scales, alleviating the feature forgetting problem caused by inconsistent multi-scale features. Thirdly, a lightweight shared convolution detection head (LSCD) is introduced, utilizing cross-scale parameter sharing and group normalization strategies to significantly reduce the computational overhead of the detection head while enhancing the detection accuracy of small targets. Finally, the proposed method is validated on a self-built multi-target defect dataset of transmission lines. Experimental results show that PFD-YOLO maintains a high inference speed while improving mAP50 and mAP50.95 to 86.6% and 53.9%, respectively, representing an increase of 3.9% and 3.3% compared to the original YOLOv11n model. This effectively enhances the detection accuracy and robustness of transmission line defects and foreign objects, demonstrating good engineering application value.

Abstract:Aiming at the key challenges in steel surface defect detection, such as insufficient accuracy, high missed detection rate of small-sized defects and bounding box localization deviations, this paper proposes an improved steel surface defect detection model YOLOv8-CEW （closed eyes in the wild） based on YOLOv8. The model achieves performance breakthroughs in feature extraction, feature fusion, and bounding box regression through three core improvements: The integration of the CFBlock （ConvFormer block） module, the efficient up convolution block (EUCB) and the WIoU loss function. Specifically, the introduced CFBlock module is fused with the existing C2f module in the Backbone and replaces the internal bottleneck to optimize the feature capture mechanism for the unique morphologies of steel defects (slender and winding, dense and discrete), enhancing the discriminative ability of multi-scale defect features; the EUCB module replaces traditional upsampling to adapt to the drastic scale changes of defects under different shooting distances, optimize defect feature reconstruction and improve detection integrity; and the WIoU loss function dynamically adjusts the localization penalty strategy according to defect scale, strengthening the accuracy of bounding box localization. Experimental results on the NEU-DET （Northeastern University defect detection dataset） demonstrate that compared with the baseline YOLOv8 model, the improved model achieves increases of 3.2%, 1.4%, 1.1% and 0.4% in precision, recall, mAP@0.5 and mAP@0.5:0.95, respectively. Furthermore, its detection speed meets the requirements of industrial applications, indicating promising practical prospects.

Abstract:Hot-bar soldering (pulse reflow soldering) is widely used in the automated welding of micro-scale electronic components, where high requirements are imposed on both the temperature ramp rate and temperature control accuracy. Conventional control methods establish process models based on time-domain assumptions or equivalent analyses; however, model mismatch may result in temperature overshoot and degraded system stability. In this study, parameter identification is carried out based on nonlinear least-squares estimation, and an extended state-space model for temperature control of the hot-bar soldering process is established. An model predictive control （MPC） controller is designed, and the closed-loop simulation of the hot-bar soldering temperature control system is performed. Simulation results indicate that, compared with the conventional proportional-integral-derivative（PID） and Smith-PID controllers, the MPC controller exhibits stronger robustness under model mismatch and achieves superior disturbance rejection performance under external perturbations. Micro-cable welding experiments demonstrate that the MPC controller exhibits good temperature regulation performance, achieving a temperature rise rate of up to 240℃/s while significantly suppressing temperature overshoot. The temperature control accuracy during the welding process reaches ±2℃. After replacing welding heads with different structural sizes, the temperature control performance remains essentially unaffected, achieving precise temperature control of the hot-bar soldering process.

Abstract:Underwater submersibles operating in emergency conditions are susceptible to ferromagnetic interference from the mother ship, leading to degraded positioning accuracy. To address this issue, a positioning error correction method for underwater submersibles based on reverse attitude update is proposed. When initial magnetic interference exists, the method first performs inertial navigation and stores the raw data. After the submersible detaches from the mother ship and the magnetic signal returns to normal, it recursively recalculates and corrects the original initial attitude angle using the inertial-phase data and reverse attitude update algorithm. Subsequently, inertial navigation and integrated navigation are re-executed with the revised initial attitude angle to eliminate the systematic errors induced by the original initial attitude angle. This method enables the estimation and compensation of both initial and process errors by leveraging the later-stage integrated navigation results. Experimental results demonstrate that, compared with the traditional integrated navigation method, after compensating for the initial heading and horizontal attitude errors of the submersible, the average position error during navigation is reduced by 87.8% and the average velocity error by 91.3%, which significantly improves the positioning performance of the integrated navigation system for underwater submersibles in emergency scenarios.

Abstract:As pivotal technologies in non-cooperative communications, signal parameter estimation and modulation recognition play a decisive role in electromagnetic signal analysis and intelligence extraction. Confronted with practical challenges including unknown signal parameters, limited model generalizability and costly post-deployment updates, this paper proposes and implements a deep learning-based blind identification method for digital modulation signals. The proposed method innovatively designs a feedforward signal processing framework that tightly integrates symbol rate estimation, adaptive sampling rate control and intelligent modulation identification. This integrated framework breaks away from the limitations of conventional paradigms that treat these functions as isolated, sequential operations, thus yielding a highly efficient and universally applicable signal processing solution. Experimental results and practical validation confirm our approach substantially reduces training expenditures while simultaneously improving generalization capacity and robustness to unknown parameters, showing strong potential for engineering applications.

Abstract:To address the issues of large collision detection scale and low collision detection efficiency in complex urban obstacle environments, which result in slow solving speeds for traditional visibility methods, a hierarchical convex hull visibility method is proposed. By pre-screening polygons in the environment using the principal convex hull, it narrows the scope of obstacles involved in visibility detection, reducing detection time. Simultaneously, a local convex hull visibility detection method is introduced. It performs polygon pair traversal to construct local convex hulls and employs bridge points for regional segmentation. This approach avoids invalid detections inherent in traditional global pair traversal, substantially improving detection efficiency. Urban environments of varying scales are constructed using real city data. Simulation comparisons are conducted against baseline path planning algorithms, related improved algorithms and existing visibility methods in the literature. Results demonstrate that the hierarchical convex hull visibility graph method rapidly generates high-quality paths, reducing planning time by 99.66%, 94.38% and 94.06% compared to traditional visibility graph methods, bidirectional rapidly-exploring random trees （Bi-RRT*） and Informed-RRT*, respectively. The proposed method effectively addresses the computational efficiency challenge for UAV path planning in complex urban environments.

Abstract:Aiming at the problems of low accuracy and long inference time in bullet hole detection technology for chest ring targets in shooting sports, a lightweight algorithm based on YOLOv8 is proposed. Firstly, a MobileNetV4 backbone enhanced with a convolutional block attention module (CBAM) is designed to enhance the model′s accuracy while minimizing its computational complexity and the number of parameters. Secondly, a C2fGhost-AFPN-P2 (asymptotic feature pyramid network) feature aggregation network is proposed, which increases the network′s perception of small targets by combining the C2fGhost module, AFPN and P2 output layer. Finally, the loss function of the algorithm is set to WIoU to weaken the impact of the imbalance in the number of positive and negative samples and improve the regression accuracy of the prediction box. The experimental results on the self-made dataset show that the mAP50 is 0.973, which is 16.0% higher than the original algorithm; the number of parameters is 2.43×106, which is 19.3% lower than the original algorithm; and the GFLOP is 8.0, which is 2.4% lower than the original algorithm. The experimental results prove that M-A-YOLO has better performance in bullet hole detection.

Abstract:In response to the problem of the denial-of-service (DoS) attacks that the network layer of cyber physical power system (CPPS) is prone to, a DoS attack detection and defense method integrating state estimation and model predictive control (MPC) is proposed. On the sensor side, a multi fading factors adaptive Kalman filter is adopted for state estimate. This algorithm overcomes the drawbacks of the traditional Kalman filter which is prone to divergence and the single fading factor adaptive Kalman filter which has low estimation accuracy, and a residual detection algorithm is designed based on the estimation results to realize attack detection. On the controller side, MPC is introduced, which calculates the control signals for the future period by using the newly received state estimation signals to compensate for the data loss caused by DoS attacks. Meanwhile the predictive control sequence is designed on the basis of Lyapunov stability theory and linear matrix inequalities (LMI), which solves the problem of system instability caused by data loss. Finally, simulation verifications are conducted on the IEEE 9-bus and IEEE 30-bus test systems. The results show that the proposed detection method can detect DoS attacks in a timely manner, and the defense method can effectively maintain the stable operation of CPPS under DoS attacks.

Abstract:As a core component of reusable thermal protection systems for spacecraft, ceramic fiber heat shields rely on the integrity of their borosilicate glass coatings to determine aerodynamic performance and operational safety. However, traditional 3D point cloud detection methods face challenges in defect localization and measurement due to the high similarity between complex surface textures and actual defect morphologies during inspection. To address these issues, this study proposes a two-stage measurement protocol. This method first performs 2D defect localization. Initial candidate area screening using relaxed threshold values on depth maps, followed by relative slope feature extraction to refine detection and eliminate texture interference; it then conducts 3D local measurement. Mapping localized defect regions to 3D point clouds and applying RANSAC/DBSCAN algorithms for precise dimensional analysis. Experimental results demonstrate significant improvements: compared to edge detection methods, the proposed approach achieves 83.3% defect detection accuracy (from 50% baseline) with the false positive rate reduced to 33.33%. When compared to global RANSAC/DBSCAN methods, it improves detection rate from 50% to 83.3% while reducing false positives from 50% to 33.33%, with detection capability extending to minimum 30 μm depth defects. The framework′s superior localization performance under complex texture interference is conclusively validated.

Abstract:In response to the problems existing in the current research on composite damage identification, such as high feature dimensions, excessive redundant information and difficulty in distinguishing multiple damage modes, a damage identification method based on principal component analysis and Hilbert marginal spectral energy features for acoustic emission is proposed. Acoustic emission signals of epoxy resin, carbon fiber bundles, and composite were obtained through tensile experiments. Principal component analysis was used to reduce the dimension of 6 types of feature parameters and extract the principal components of high-energy fracture and frequency domain features; the Hilbert-Huang Transform was used to obtain the marginal spectrum and establish a quantitative criterion based on the concentration of band energy. The results show that the cumulative contribution rates of the first three principal components for the three types of specimens are 84.7%, 88.25% and 86.6%, respectively. These effectively extract two main damage modes: high-energy fracture (PC1) and frequency domain features (PC2). Further analysis of waveform energy using Hilbert-Huang Transform reveals that the damage energy of pure resin is concentrated in the 100~150 kHz frequency band (accounting for approximately 60%), the damage of the fiber bundle shows a three-stage evolution in the F4~F6 mid-frequency band, and the energy distribution of the composite is highly dispersed (with a dispersion mode accounting for 71%). This method can effectively distinguish matrix cracking, fiber fracture, and interface debonding, providing a quantitative identification basis for composite damage monitoring.

Abstract:To effectively suppress noise in intravoxel incoherent motion (IVIM) images while preserving structural information without the need for clean labels or repetitive acquisitions, this paper proposes a self-supervised denoising model based on similarity-constrained masking. Built upon the blind-spot denoising framework, the model learns from a single noisy image by integrating image neighborhood similarity with an adaptive masking strategy. In simulation experiments, at a signal-to-noise ratio(SNR) of 30, the proposed method reduces the signal mean square error (MSE) by 36.92%. Furthermore, the root mean square error (RMSE) of the parameters F, D and D_star decreases by 4.39%, 11.43% and 3.29%, respectively, while their structural similarity index measure (SSIM) increases by 2.73%, 0.21% and 3.37%. In real-world experiments, the SNR improvement ratios in gray matter (GM) and white matter (WM) regions increase by 0.40% and 3.85%, respectively. Clinical case analysis further demonstrates that the contrast-to-noise ratio of D and D_star values in the meningeal region improves from 0.064 and 0.005 to 0.271 and 0.026, respectively. Experimental results indicate that this method consistently achieves the highest and most stable contrast-to-noise ratio levels across all parameters. The proposed model effectively reduces noise while preserving tissue structure and parametric information, providing an efficient self-supervised solution for high-quality IVIM imaging and its clinical applications.

Abstract:To improve the recognition accuracy of small maritime targets in remote sensing images, a YOLO algorithm integrating multi-kernel sub-pixel enhancement and contextual attention is proposed based on the YOLO11n object detection framework. This method simultaneously employs 3×3 and 9×9 convolutional kernels to extract features with different receptive fields and adopts a super-resolution reconstruction approach to design a sub-pixel enhancement module that captures more detailed information, thereby enhancing the feature extraction capability of the backbone network. On the basis of the original C3K2 structure, a C3K2_context aggregation attention (C3K2_CAA） module is designed to capture richer contextual information, assisting the model in more accurately recognizing targets and improving detection performance in complex scenarios. The network′s detection effectiveness is validated using the publicly available SeaShips dataset. The results show that double-core sub-pixel enhancement module (DSPE）-YOLO algorithm achieves AP, mAP50 and mAP50-95 of 97.2%, 99.5% and 71.6%, respectively. Compared with the baseline YOLO11n network, the recognition accuracy increases by 4.3%, 9.4% and 4.4%, respectively. For ship detection, the AP is improved by 0.4%, 7.7%, 4.7% and 8.7% compared with PP-YOLOE-s, YOLOV6, YOLOV8-n and YOLOV10-n demonstrating excellent object detection performance.

Abstract:In recent years, unmanned aerial vehicles have been widely applied in the field of remote sensing image acquisition. However, in the task of target detection within remote sensing images, single-modal visible light images exhibit certain limitations, as their detection performance significantly degrades in dark environments or under adverse weather conditions such as haze. To address this issue, this paper proposes a dual-channel remote sensing image target detection network based on YOLO, named YOLO with dual-channel fusion and Mamba（YOLO-DFM）. This network integrates the Mamba architecture with the YOLO model and incorporates the dehazing module C2f with dynamic-range histogram self-attention (C2f_DHSA） to enhance the detection capability of visible light images. Additionally, we introduce a visible-infrared profound semantic fusion module (PSFM), which effectively fuses visible and infrared images, leveraging the complementary advantages of both modalities to improve the robustness and accuracy of remote sensing image target detection. For feature extraction and fusion, we propose a dual-convolution （feature multi-scale composite convolution，FMCConv） module, which combines the advantages of space-to-depth convolution (SPDConv） and deep robust feature downsampling (DRFD）. By employing an innovative feature fusion strategy and computational optimization techniques, this module enhances target detection accuracy in complex multi-scale backgrounds and is integrated into the YOLO-DFM network. We validate our proposed method on the vehicle detection in aerial imagery (VEDAI）, low-light visible-infrared paired (LLVIP） and DroneVehicle datasets. Experimental results demonstrate that, compared to the original YOLOv8, YOLO-DFM improves the mAP metric by 6.8%, 1.2% and 1.6%, respectively, significantly enhancing detection performance and confirming the model′s effectiveness. In conclusion, the YOLO-DFM network proposed in this paper provides an efficient and reliable solution for multi-modal remote sensing image target detection.

Abstract:Underwater concrete structures in hydraulic dams are prone to defects. Traditional manual inspection methods are inefficient and carry high risk factors. Existing underwater 3D reconstruction techniques struggle to achieve high precision due to poor image quality and complex aquatic environments. Therefore, for underwater structural defect detection, this paper proposes a 3D reconstruction solution integrating a multi-scale fusion image enhancement algorithm with an underwater refraction model. This approach combines adaptive color correction and an improved Retinex algorithm through non-subsampled shearlet transform (NSST） algorithm-based multi-scale fusion to optimize image quality, then utilizes the underwater refraction model to complete 3D reconstruction. Experiments demonstrate that this method outperforms traditional underwater image enhancement models in terms of information entropy (IE), patch-based contrast quality index (PCQI) and underwater color image quality evaluation (UCIQE) metrics.The processed images achieve an optimal IE value (7.803 6), with actual reconstruction results showing optimal detail textures and edge contours. This approach effectively enhances the reconstruction accuracy of underwater structural defects, providing reliable support for underwater inspection.

Abstract:Medical image segmentation is one of the core tasks in medical image analysis and plays a vital role in clinical diagnosis, surgical planning, and treatment evaluation. To address the limitations of existing methods in global context modeling, local detail preservation and computational efficiency, this paper proposes a novel medical image segmentation method based on a multi-scale visual state space U-Net. This method integrates a state space modeling mechanism, a multi-scale parallel wavelet convolution module and an improved large-kernel grouped attention gating module, aiming to enhance cross-layer feature fusion capability and detailed feature representation. Furthermore, this paper introduces a hierarchical cascaded prediction strategy, which progressively generates and fuses prediction maps at multiple semantic levels of the decoder, thereby further improving the model′s robustness and adaptability to multi-scale targets. Experimental results on two public datasets, Synapse and automated cardiac diagnosis challenge (ACDC）, demonstrate that the overall performance of the proposed method surpasses existing mainstream methods, validating its effectiveness and superiority.

Abstract:To address the issue of complex non-uniform blurring in images captured by drones during high-speed flight and attitude jitter in dynamic road inspections, which severely degrades image quality and impacts subsequent object detection, this paper proposes an improved DeblurGAN-V2 image deblurring algorithm guided by inertial sensors. This algorithm utilizes camera inertial measurement unit (IMU) data as physical constraints, designs a gradient direction consistency loss function to enforce motion trajectory recovery, and employs dual path-aware loss to balance texture and structure. Additionally, an adaptive weighting discriminator is designed to evaluate the authenticity of generated images at different scales, thereby enhancing image quality. Experiments on a self-built drone road inspection dataset demonstrate that compared to the baseline DeblurGAN-v2 model, peak signal-to-noise ratio (PSNR） and structural similar ityindex measure (SSIM） improve by 1.52 dB and 0.056, respectively. Furthermore, in downstream vehicle object detection tasks, the false positive rate and false negative rate decrease by 9.2% and 18.1%, respectively.

Abstract:To address the limitations of unmanned aerial vehicle vision-and-language navigation (UAV-VLN) in multimodal feature fusion, global environment perception and long-distance planning, a history-enhanced dual-scale mixture-of-experts network is proposed. The proposed framework constructs a dual-scale structure composed of a global mixture-of-experts network and a local mixture-of-experts network. Specifically, the global network integrates linguistic descriptions with navigational map features to perform global semantic reasoning, while the local network combines visual observations and linguistic features to enable fine-grained semantic perception, thereby enhancing cross-modal understanding at different scales. To strengthen temporal modeling and long-term dependency capture, a historical memory encoder based on a matrix long short-term memory (mLSTM) network is introduced to dynamically fuse historical observations with the current state. Furthermore, a goal-progress co-prediction module is designed to jointly estimate the agent′s spatial position and task stage through a goal prediction head and a progress prediction head, thereby improving the continuity and interpretability of navigation decisions. Experiments conducted on the city-scale UAV-VLN CityNav dataset demonstrate that the proposed method achieves a success rate (SR） of 16.42% and a success weighted by path length (SPL） of 18.26%, outperforming the baseline model by 10.04% and 12.18%, respectively. These results verify that the proposed approach substantially enhances cross-modal understanding, temporal reasoning and long-distance navigation capability under urban scenarios.

Abstract:Vehicle re-identification (Re-ID) aims to accurately identify and associate the same vehicle across different camera views to achieve cross-regional, continuous tracking and trajectory analysis. However, existing Re-ID methods often suffer from insufficient feature representation and poor robustness in complex scenarios due to camera viewpoint variations, occlusions and high intra-class similarity. To address these challenges, this study proposes a dual-branch network fusing regional features with dynamic side information. This study design a regional contextual association module (RCAM) using multi-head attention to adaptively refine key local features, and a dynamic side information embedding (D-SIE) mechanism to flexibly incorporate camera and viewpoint data. By employing CutMix augmentation, the model′s robustness and generalization are significantly enhanced. On VeRi-776, the method achieves 83.6% mAP and 97.5% Rank-1; on the 800 subset of the VehicleID dataset, the Rank-1 accuracy reaches 85.4% and the Rank-5 accuracy hits 97.8%. Cross-domain experiments also verify the strong generalization performance of the model. Experimental results demonstrate that the model proposed in this paper significantly improves the model robustness and fine-grained feature discrimination capability for the vehicle re-identification task.