Abstract:To address the performance distortion issue in the light-load region faced by the state-trajectory-model-based synchronous rectification method for CLLC resonant converters, this paper proposes an optimized synchronous rectification strategy based on the state trajectory model. Through an in-depth analysis of the DCM of the converter under light-load conditions, the boundary conditions for current DCM are derived, enabling accurate identification of the current discontinuous mode and effectively suppressing circulating current losses. The proposed method requires no additional detection devices and relies solely on basic input and output electrical parameters to achieve control, significantly reducing hardware complexity and cost. Experimental validation on an 800 W platform demonstrates that the proposed scheme achieves a peak conversion efficiency of 97.38% across the entire operating frequency range, with notable improvements in dynamic response and overall efficiency compared to traditional methods.

Abstract:To enhance the development efficiency and real-time signal processing capabilities of flexible sensors, this paper presents an integrated signal processing platform specifically designed for resistive flexible sensors. The platform enables rapid acquisition and continuous processing of sensor signals while addressing issues such as nonlinear output and poor stability. It consists of a resistance acquisition system and a signal processing system: the acquisition module, based on an STM32 microcontroller, supports real-time acquisition from eight channels of flexible sensors; the processing module integrates a segmented linear mapping algorithm based on calibration points and supports 11 filtering methods—including hybrid adaptive filtering—as well as real-time waveform visualization. Experimental results demonstrate that the platform achieves precise acquisition of flexible sensor signals, significantly improves signal linearity and output stability, and supports fully integrated real-time operation across the acquisition and processing stages. This platform provides a reliable data foundation and methodological support for the development and application of flexible sensors.

Abstract:To address the limitations of traditional data acquisition systems in air-to-air missile testing, such as insufficient multi-channel synchronous transmission, high-speed storage, and anti-interference capabilities, this study designs an anti-interference real-time data acquisition system based on the ZYNQ-7000 series and an EMMC 5-1-compliant storage module. The system enhances signal robustness through isolated LVDS interfaces and RS422 transceivers, achieves synchronous multi-source data acquisition via a tunable multi-channel sampling rate/baud rate architecture and hybrid framing technology, and ensures timing stability for high-speed EMMC read/write operations through dynamic calibration of data sampling points. Additionally, DDR3 cache is integrated to optimize burst data processing. Experimental results demonstrate that the single EMMC system achieves a write speed of 157.6 MB/s, a read speed of 180.1 MB/s, and a host computer data transmission rate of 40.78 Mb/s. By synergizing hardware isolation, framing optimization, and high-speed storage design, significantly enhancing real-time performance and anti-interference capabilities for multi-source data acquisition in high-dynamic environments. This system provides high-reliability data support for evaluating air-to-air missile performance.

Abstract:The dynamic marine environment induces platform motions that compromise the reliability of ocean observation platforms and measurement instruments, leading to errors in the measurement of ocean dynamic parameters such as wind, waves and currents. Throughout its deployment and operational cycle, an ocean observation platform experiences irregular motions. Conventional attitude estimation algorithms based on a nine-axis Micro Inertial Measurement Unit (MIMU), such as Kalman filtering or complementary filtering, often fail to achieve timely convergence, resulting in significant errors in attitude computation. To address this issue, this study proposes a two-stage extended Kalman filtering system for MIMU-based attitude estimation. In the first stage, the system utilizes acceleration data to estimate the initial attitude angles, thereby improving filter convergence speed. In the second stage, it employs hierarchical processing of roll, pitch, and yaw angles to enhance system robustness. Laboratory and field experiments at sea demonstrate that the proposed method achieves rapid convergence and high-accuracy attitude estimation for ocean observation platforms operating under non-stationary motion conditions.

Abstract:Aiming at the lack of multi-step prediction methods for soil landslide displacement and the issue of significant prediction errors over extended time horizons, this paper proposes a multi-step landslide displacement prediction method based on a parallel model with a multi-level attention mechanism. The method employs a multi-input multi-output prediction strategy. Utilizing a Transformer encoder branch incorporating a multi-head attention mechanism and a bidirectional gated recurrent unit (BiGRU) branch optimized with a global attention mechanism (GAM), the two parallel network branches process historical landslide monitoring data. The landslide feature information extracted by the parallel networks is then fused via a cross attention mechanism (CAM), subsequently outputting the predicted multi-step displacement values. Experimental results demonstrate that the multi-level attention mechanism model achieves a mean absolute error (MAE) of 2.17 mm, a root mean square error (RMSE) of 3.05 mm, and a coefficient of determination (R2) of 0.968 9 in multi-step landslide displacement prediction. Compared to other models, it yields the lowest errors and the optimal R2 result. The model exhibits more stable prediction performance over long time horizons, facilitating the early anticipation of landslide development trends. This provides crucial technical support for landslide prevention and mitigation.

Abstract:In Vehicle-to-Everything scenarios, the mobility of devices and the complexity of the environment render them more vulnerable to malicious attacks, necessitating a secure and efficient authentication mechanism. Radio Frequency Fingerprinting (RFF) offers a novel approach to identity authentication in V2X networks. However, as device fingerprints are extracted directly from wireless signals, their stability is highly susceptible to channel variations. The combined effects of the wireless channel and receiver noise cause distortion in the received signal, making it challenging to directly isolate the authentic features of the transmitted signal. To address these issues, this paper proposes an RFF extraction method based on an improved Linear Minimum Mean Square Error channel estimation for the Physical Sidelink Broadcast Channel. First, a channel estimator based on the LMMSE criterion is constructed. By exploiting the time-frequency two-dimensional statistical properties of the channel, a corresponding 2D correlation matrix is established, which effectively captures the intrinsic coupling relationship between time-selective and frequency-selective fading. Based on this matrix, the random channel response can be optimally separated from the received signal. Subsequently, a channel equalization operation is performed to recover the hardware fingerprints contaminated by the channel, restoring their original feature information. Finally, a structurally optimized dual-branch heterogeneous neural network is employed for deep representation learning and high-precision classification of these hardware fingerprints. Experimental results demonstrate that, under low signal-to-noise ratio conditions, the proposed method achieves classification accuracy of 95.46% and 92.05% in static and mobile scenarios, respectively.

Abstract:To address the navigation accuracy challenges of low-cost unmanned vehicles in complex motion environments, this paper proposes an integrated navigation filtering algorithm based on multi-modal motion characteristic decomposition. The method combines Kalman Filter and Cubature Kalman Filter dynamically selecting the optimal filtering strategy according to the motion characteristics of the vehicle. In low-dynamic environments, the Kalman Filter is used to improve computational efficiency, while in medium-dynamic environments, the Cubature Kalman Filter is applied to enhance nonlinear state estimation capabilities. The proposed method is validated through simulations using the Precise Strapdown Inertial Navigation System toolbox, analyzing UAV and UGV trajectories. Experimental results show that compared to traditional filtering methods, the proposed algorithm reduces position estimation errors by 25% in UAV scenarios and improves computational efficiency by 50% in UGV scenarios.

Abstract:With the diversification of network attack means and the complication of traffic characteristics, the detection of network malicious traffic is facing increasingly severe challenges. Traditional traffic detection methods gradually fail to meet the needs of modern network environments in terms of accuracy and reliability, especially in the case of high-dimensional data and complex attack patterns. To address the above issues, this paper proposes a network malicious traffic detection model based on the Crested Porcupine Optimization Algorithm, Bidirectional Long Short-Term Memory Network, and Kolmogorov-Arnold Network. The model uses the Bidirectional Long Short-Term Memory Network to capture the bidirectional temporal features of traffic data, combines the nonlinear mapping of the Kolmogorov-Arnold Network to enhance feature expression capabilities, and optimizes hyperparameters through the Crested Porcupine Optimization Algorithm to improve model performance. Experiments are conducted using the CIC UNSW-NB15 enhanced dataset. The experimental results show that the model achieves accuracies of 99.12% and 94.15% in binary classification and multi-classification tasks, respectively, significantly outperforming other models. In addition, when dealing with class imbalance, the model particularly enhances the detection capability for minority class samples such as Backdoor and Worms.

Abstract:Aiming at the three problems of complex background, inconsistent target size and small proportion of defective areas to be inspected in aerial insulator images taken by UAVs during transmission line inspection, a lightweight insulator defect detection algorithm, MHD-YOLO, is proposed. Firstly, a feature extraction network MAFNet is introduced into the backbone network of YOLOv8, and hybrid convolution is used to enhance the feature extraction capability of the network under complex background. Second, a feature fusion network, HS-FPN, is used to realize feature fusion at different scales, and combined with a lightweight dynamic up-sampling method, DySample, to improve the quality and efficiency of up-sampling. Then, a lightweight detection head CSH is designed, which significantly reduces the number of parameters in the detection layer and the computation amount by using the shared convolution method. Finally, the NWD loss function is introduced to improve the localization accuracy of the model for small targets. The experimental results demonstrate that the MHD-YOLO target detection algorithm reduces the number of parameters by 43.8% compared with YOLOv8, and improves the detection accuracy by 5.1% on the insulator defect detection dataset. The improved algorithm is significantly more effective in detecting insulator defects, and the model complexity is greatly reduced, providing a more effective method for deployment on embedded devices.

Abstract:To address the issue of performance degradation in rolling bearing fault diagnosis under varying operating conditions caused by distribution discrepancies between the source and target domains, this paper proposes a novel fault diagnosis method that integrates a multi-head self-attention mechanism with dynamic joint distribution adaptation. Firstly, a multi-head self-attention mechanism is incorporated into the feature extraction module to extract more discriminative and domain-invariant features from raw vibration signals. Secondly, maximum mean discrepancy and local maximum mean discrepancy are employed to align the marginal and conditional distributions, thereby reducing the distribution difference between the source and target domains. Finally, a dynamic weighting factor is designed to adaptively adjust the importance of marginal and conditional distribution alignment, enhancing the robustness and generalization ability of cross-domain fault diagnosis. The experimental results demonstrate that the proposed method achieved classification accuracies of 99.84% and 98.97% on two public datasets, significantly outperforming other approaches. Moreover, it maintained strong stability and robustness under severe noise interference, providing an effective solution for rolling bearing fault diagnosis.

Abstract:Ankylosing spondylitis is a chronic inflammatory disease whose early diagnosis depends on the accurate identification of pathological features in the sacroiliac joint. However, due to the complex anatomical structure of the sacroiliac joint, the multiscale heterogeneity of lesions, as well as interference from partial volume effects and noise in CT imaging, the accuracy of traditional segmentation methods often fails to meet clinical demands. To address these challenges, this study proposes a Multiscale Attention-Guided U-Net (MAG-UNet). The model enhances local-global feature representation through a Multiscale Feature Fusion (MFF) module, integrates spatial-channel adaptive weighting via a Dual-path Attention (DA) mechanism, and introduces a Large-kernel Grouped Attention Gate (LGAG) to resolve cross-scale feature coupling issues. Experiments conducted on a dataset provided by Shanxi Bethune Hospital demonstrate that MAG-UNet achieves significant performance improvements in sacroiliac joint CT segmentation, with a Dice coefficient of 92.4% and an Intersection over Union (IoU) of 86.0%, surpassing the baseline U-Net model by 3.4% in IoU. This study provides a reliable technical solution for the early diagnosis of AS, offering substantial clinical value and broad potential for practical application.

Abstract:Flexible DC transmission systems based on modular multilevel converters (MMC) typically use traditional sliding mode observation techniques to observe state variables, which can significantly reduce the accuracy of system state reconstruction and transient response speed under sudden changes in operating conditions. A discrete-time logarithmic sliding mode observation technique (DTLSMO) is proposed, which constructs dynamic boundary layer functions of observation error norm and expected amplitude, achieves adaptive nonlinear mapping of observer gain, and avoids the parameter tuning process that relies on empirical trial and error. Establish a discretization model of MMC under parameter mismatch and analyze its observability. Design a logarithmic sliding mode approaching law based on active damping characteristics and inject it into the bridge arm dynamic equation. Use the second harmonic circulating current injection algorithm to achieve the coordinated optimization of capacitor voltage balance control and circulating current suppression. Prove the asymptotic stability and generalized convergence of the proposed observer. Analyze the DC voltage 9.6 kV MMC simulation system on MATLAB/Simulink platform and build a hardware prototype for experimental verification. Introduce evaluation indicators such as ITAE to compare and analyze the dynamic performance differences between DTLSMO and traditional time discrete sliding mode observer (DTSMO) and adaptive observer (AO). It is proved that the proposed observer improves the observation accuracy by 0.935% and 2.535% respectively when the bridge arm inductance is mismatched by 20%. It can still demonstrate its fast response ability under uncertain capacitor values and sudden changes in DC voltage, and has good robustness.

Abstract:To address the inevitable limitations of current LiDAR-only 3D detection methods, which are affected by point cloud sparsity—where LiDAR-scanned point clouds exhibit significantly higher sparsity at long range compared to short range, leading to imbalanced positive and negative samples during model training—we propose a novel multi-modal framework named MCA-VoxelNet, based on pseudo-point-cloud fusion.It consists of two key designs: the pseudo-point clouds generated by depth completion are utilized to solve the problem of point cloud sparsity, and a large number of nearby redundant voxels are discarded through the distance-aware sampling module to enhance computational efficiency; a multi-stage cascaded attention detection structure is employed to aggregate the target features of multiple detection stages, balance the number of positive and negative samples, and gradually improve the region proposals output by the Region Proposal Network. Experiments on the authoritative KITTI autonomous driving dataset demonstrate that MCA-VoxelNet achieves an inference speed of 17.54 FPS and attains car detection accuracies of 94.19%, 85.93%, and 86.17% on the easy, moderate, and hard difficulty levels, respectively. These results outperform the second-best method by 2.64%, 1.16%, and 1.91%.

Abstract:The deadbeat predictive current control for permanent magnet synchronous motor features fast dynamic response, excellent steady-state performance, and easy digital implementation. However, traditional deadbeat predictive current control exhibits significant steady-state current errors when motor parameters are mismatched. Therefore, this paper proposes an adaptive deadbeat predictive current control method for permanent magnet synchronous motor, which reduces the impact of parameter mismatches on current control performance. Firstly, the proposed method uniformly treats resistance, inductance, and flux linkage parameter perturbations as disturbances and designs adaptive disturbance laws for the dq-axis by incorporating adaptive control. On the basis, the disturbance is estimated online through the disturbance adaptive law, and disturbance compensation is carried out in the traditional deadbeat predictive current controller. This approach effectively reduces the adverse effects of mismatched parameters on the traditional deadbeat predictive current control. Secondly, the stability of the proposed control method was proven using discrete Lyapunov stability theory, and the values of the disturbance adaptive gains were determined through classical control theory. Finally, simulation and experimental results demonstrate that the designed disturbance adaptive law can rapidly and accurately estimate disturbances. Compared with traditional deadbeat predictive current control, the proposed method effectively suppresses the impact of mismatched parameters on current dynamic performance and steady-state error，while reducing the total harmonic distortion of the current.

Abstract:In the task of small target defect detection on steel cables, there are common problems such as low detection accuracy, high missed detection rate and frequent false detection, which are particularly obvious in the detection scenario with more small sizes. The main reasons for such problems include: insufficient feature extraction capability of traditional detection algorithms, lack of effective multi-scale information fusion mechanism, and insensitivity of existing loss functions to small targets. To address the above problems, a steel cable defect detection method based on improved RT-DETR is proposed. The BasicStar feature extraction module was designed in the backbone network to improve the semantic representation ability of the model in high-dimensional space; at the same time, a new multi-scale feature fusion strategy small object pyramid network(SOPN) is designed to strengthen the attention and expression ability of small targets; in terms of loss function, a focal enhancement Focaler-SIoU loss function is proposed to improve the positioning accuracy of small targets and the stability of training convergence. Experimental results on the steel cable defect dataset show that the improved model improves the average detection accuracy mAP50 by 2.1% compared with the original RT-DETR. The comprehensive performance is better than the existing mainstream target detection algorithms, which verifies the effectiveness and practicality of the proposed method for small target defect detection tasks in industrial scenarios.

Abstract:Aiming at the existing multivariate long time series prediction model in the medium and long term prediction of photovoltaic (PV) power, which has the problem of poor prediction results due to insufficient feature extraction, a multivariate long time series prediction model FFTEMixer based on learning in both frequency and time domains is proposed, which is capable of accurately predicting the PV power while maintaining a high operational efficiency. The model first uses the fast Fourier transform to project time-series data into the frequency domain. It then selectively enhances or suppresses specific frequency components through learnable frequency filters to extract global features and inter-variable correlation features. Next, an interactive convolution module is used to learn local dependencies, further enhancing feature expression capabilities. Subsequently, a feature fusion module is employed to further integrate periodic features, and establishes associations between feature variables and time stamp covariates. Finally, a multi-head self-attention mechanism is employed to comprehensively model the long-term dependencies and temporal dependencies of the sequence, thereby achieving comprehensive feature extraction from time-series data. Experimental results show that on two publicly available photovoltaic power generation datasets, the model′s predictive performance significantly outperforms the baseline model, with mean squared error (MSE) and mean absolute error (MAE) consistently achieving the lowest values. Compared to the current mainstream second-best model, its MSE and MAE are reduced by 12.6% and 15.8%, respectively, validating the model′s effectiveness.

Abstract:In light of the significance of wind power within the energy landscape and the challenges posed by its intermittency, this paper proposes an end-to-end, ultra-short-term wind power multi-step prediction model that integrates outlier processing and multi-scale feature fusion. The objective is to enhance the accuracy and stability of ultra-short-term wind power predictions, thereby providing robust support for the reliability of power system scheduling and operation. First, the RobustTSF method is employed to address time series anomalies, providing a strong assurance of the prediction model′s robustness and minimizing the disparity between abnormal time series prediction and noise label learning. Secondly, the integration of the spatial pyramid matching mapping strategy, Levy flight strategy, and adaptive T-distribution mutation strategy enhances the dung beetle optimization algorithm, significantly improving its global search capability and convergence efficiency. Meanwhile, the multi-strategy dung beetle optimization algorithm is utilized to optimize the hyperparameters of the enhanced TimeMixer model, resulting in optimal model performance. Finally, the CATimeMixer model is employed to achieve the fusion and prediction of multi-scale seasonal features and trend features. The experimental results indicate that the MAE, RMSE, and MSE decreased by 49.71%, 41.26%, and 65.50%, respectively, compared to the benchmark model multilayer perceptron, while the R2 value increased by 4.49%. This demonstrates a significant reduction in prediction error and offers a novel approach for the accurate prediction of ultra-short-term wind power.

Abstract:This study proposes a new paradigm of dynamic prior modulation based on degradation kernel decoupling evaluation to address the key challenges of kernel estimation bias and non blind method prior mismatch in blind super resolution reconstruction. By establishing a decoupling evaluation mechanism for degraded kernel window width amplitude, it is revealed that the estimation error of kernel window width has a decisive impact on the generalization performance of non blind reconstruction networks. Based on this, this work innovatively constructs a two-stage optimization framework: introducing a loss function relaxation constraint strategy in the kernel estimation stage, enhancing the compatibility between the estimation kernel and non blind priors by avoiding excessive loss functions affecting the accurate estimation of kernel window width; simultaneously design a dynamic kernel prior modulation network, adopting a dual path feature collaborative optimization mechanism. The sharpening feature module extracts image sharpening prior through high-frequency gradient enhancement, while the fuzzy attenuation feature module suppresses noise interference through mean filtering and extracts fuzzy attenuation prior features with regional degradation differences. The two generate degradation modulation vectors through prior modulation layers to achieve dynamic calibration of the kernel feature space. Experimental verification shows that dynamic kernel prior modulation network improves PSNR by 1.92 dB in Set5 dataset with 2×Gaussian blur scenes and 0.61 dB in BSD100 dataset with 4×strong noise scenes, significantly better than existing optimal methods. This method effectively solves the problem of kernel prior mismatch in complex degraded scenarios, providing an innovative solution for blind super-resolution reconstruction in actual complex degraded scenarios.

Abstract:In the key component detection task of UAV inspection of aerial images for transmission lines, a multimodal multi-scale target detection approach is proposed to address the challenges of accuracy degradation and high miss rates for small targets in single-modal detection methods. This approach integrates visible light and infrared images. First, the network constructs a parallel two-stream feature extraction backbone designed to simultaneously process visible light and infrared images. This design fully utilizes the rich color and texture detail information from the visible light images, along with the superior imaging stability and high contrast characteristics of the infrared images. Next, to facilitate cross-modal information interaction and complementarity, a Multimodal Feature Fusion Interactive Module (MFIFM) is developed. This module dynamically adjusts the fusion weights of features from different modalities, adaptively integrating the most discriminative information and effectively mitigating conflicts arising from modality differences. Additionally, to enhance the perception of small target components, a Hybrid Residual Multi-Scale Transformer (HRMS Transformer) module is incorporated into the dual-stream backbone. By utilizing a multi-head attention mechanism, hierarchical feature reorganization, and a residual-based strategy, the model′s ability to extract global context information is significantly strengthened. Experimental results demonstrate that the model′s mean Average Precision (mAP) at IoU thresholds of 0.50 and 0.50:0.95 improves by 5.35% and 4.48%, respectively, compared to existing single-modal methods. These findings confirm the effectiveness and applicability of multimodal fusion technology in transmission line inspection.

Abstract:High-quality deep-sea images are essential for the development of marine biology, topography and geology etc. In order to solve the problems of color distortion, image blur and low contrast in deep-sea images, we propose a deep-sea image enhancement network using improved U2-Net as the GAN generator. Firstly, the RSU module is introduced in U-Net to enhance the fusion of high-level and low-level information in the network. Secondly, the DA mechanism is introduced in the skip connection of U2-Net, which is used to enhance the interrelationship between the space and channel of the image, and extract the underwater color and texture details. Then, U2-Net with the DA mechanism, is used as the generator of GAN to enhance the realism of the image in the adversity. In addition, a new loss function with edge loss and perceived loss is reconstructed, called DS-Loss and the mapping relationship between deep-sea images and target images of U2-GAN is guided by DS-Loss from multiple perspectives. Finally, U2-GAN is compared with seven advanced underwater image enhancement algorithms on the self-built dataset DSIED. Compared with the second-place Sea-Pix-GAN, U2-Net improves by 5.6%, 3.9%, 5.2%, 16.0%, 7.1% and 2.4% in PSNR, SSIM, IE, UIQM, UCIQE, and PCQI, demonstrating better underwater image enhancement effects.

Abstract:In response to the limitations in flexibility and the insufficient level of intelligence in existing image processing systems, this study addresses key technical challenges in the dynamic reconfiguration domain of domestic FPGAs by proposing a logic-level dynamically scheduled image processing system. The system is developed based on the PH1A90 FPGA chip from Shanghai Anlogic Technology Co., Ltd., integrated with a Qt-based host platform. By employing logic-level dynamic scheduling principles, the system achieves near-Dynamic Partial Reconfiguration (DPR)-level full-link dynamic scheduling under the control of an intelligent host module. Experimental validation demonstrates successful hardware-software co-design between the Qt host and the PH1A90 FPGA, achieving four-interface heterogeneous data acquisition, dual-sensor collaborative imaging, dynamic configuration of algorithmic processing chains, and real-time execution of twenty-two dynamic ISP algorithms. The system outputs 1 080P@60 fps video streams via HDMI, validating its dynamic scheduling capabilities across interfaces, workflows, and algorithms at the logic level. Furthermore, the results confirm the industrial-grade reliability of domestic FPGA solutions, contributing to the advancement of independent and controllable technologies in fields such as intelligent security and industrial inspection.

Abstract:To address the issues of low detection accuracy and missed or false detections of small defects in aluminum profile production, this paper proposes an improved YOLOv12n-based method, termed YOLO-PCSU, for surface defect detection. First, a novel A2C2f-PConv structure is designed by integrating PartialConv into the A2C2f module of YOLOv12n, enhancing feature extraction while reducing redundant computation and memory access. Second, CoordAttention is introduced into the backbone to improve detection accuracy without increasing computational cost. Third, the SEAM attention module is added to the detection head to mitigate missed and false detections of small targets. Finally, the U-IoU loss replaces the original CIoU loss to accelerate convergence and enhance prediction precision. Experiments on an aluminum profile defect dataset demonstrate a detection accuracy of 90.3%, with a 2.3% mAP@0.5 improvement over the baseline YOLOv12n, a 9% reduction in parameters, and a 14% reduction in computation. Additional evaluations on the VOC2012 and Northeastern University hot-rolled strip steel surface defect datasets confirm the robustness of the proposed approach.

Abstract:Aiming at the feature ambiguity and misdetection problems in bird′s eye view lane line detection caused by environmental disturbances such as sudden changes in illumination and extreme weather, this paper proposes a causal interventionbased BEV lane detection framework. First, to enhance the representation of features during BEV spatial transformation, composite positional encoding is designed and fused to front view features to maintain spatial continuity and consistency. Second, the causal intervention module is constructed after acquiring the BEV features. The causal intervention module aims to explicitly decouple the lane line features from the environmental disturbances by generating counterfactual features to improve the stability of the model in extreme environments. Finally, the dynamic calibration of multi-scale features and interference suppression is accomplished by introducing the feature fusion module, and the global attention mechanism is utilized to achieve the enhancement of BEV features. The experimental results show that in the three subsets of the Apollo dataset, the F1 values are improved by 0.8%, 1%, and 3% compared to the model with the 2nd performance, and the F1 values are also optimal in the challenging scenarios within the OpenLane dataset that contain extreme weather, night, and intersections. The explicit decoupling of lane line features and environmental disturbances is successfully realized, providing a highly robust solution for autonomous driving perception in complex environments.

Abstract:To address the challenges of small and densely packed targets in drone aerial images, which are prone to missed and false detections, this paper proposes an improved multi-scale target detection model, UCM-YOLOv8, based on YOLOv8n, for complex backgrounds in drone aerial photography.Initially, a pyramid network structure that integrates aggregation and diffusion mechanisms is designed, enabling features at each scale to capture detailed contextual information. Second, a task dynamic alignment detection head is introduced to learn interactive features from multiple convolutional layers, enhancing detection precision. Furthermore, the effective integration of the convolutional additive self-attention mechanism with the C2f module further strengthens the network′s feature representation capacity. Finally, the Wise-Inner loss function is employed to replace the original CIoU loss function, suppressing harmful gradients caused by low-resolution images.The proposed model was validated through comparative and ablation experiments on the VisDrone2019 dataset. Results show a 10.8% improvement in mAP50 over the baseline model and a 9.6% reduction in parameters. These findings demonstrate the model′s superior performance in detecting small targets from drone perspectives, making it well-suited for drone aerial image applications.

Abstract:With the acceleration of urbanization, the continuous increase in domestic waste has posed a severe challenge to the ecological environment. Therefore, intelligent sorting technology based on target detection has become a key solution. Aiming at the problems of insufficient accuracy and low deployment efficiency of existing detection models in complex scenarios, an improved YOLO11 domestic waste detection model is proposed. By introducing deformable convolution and a self-designed three-branch coordinate attention mechanism, an enhanced deformable convolution module is constructed, which is used to reconstruct C3k2 in the backbone network, significantly improving the model′s feature extraction capability for targets in complex background. In addition, a content-aware feature recombination operator is adopted to replace the upsampling in the neck network, enhancing the feature reconstruction effect. An exponential moving average sliding loss function is introduced to improve detection accuracy effectively and accelerate model convergence. Experiments on the optimized Huawei Cloud domestic waste dataset show that the improved model achieves 76.5% and 64.6% in mAP@0.5 and mAP@0.5:0.95 metrics, respectively, with an increase of 1.8% and 1.7% compared to the baseline model. Compared with other mainstream detection algorithms, the improved model has a parameter count of only 2.8 M, making it more suitable for mobile deployment.