Abstract:To address the poor performance of existing search-and-rescue detection methods in complex environments, this study proposes an enhanced algorithm named DAAS-RT-DETR. Based on the real-time detection transformer (RT-DETR) framework, the method first applies dual convolution kernels (DualConv) to strengthen multi-scale feature extraction. The global attention mechanism (GAM) is then incorporated into the attention-based intrascale feature interaction (AIFI) module to jointly capture spatial and channel dependencies. An adaptive gated fusion module (AGFM) is further introduced to improve feature interaction efficiency, and the loss function is replaced with scale-based dynamic loss (SBD Loss) to achieve more accurate object localization.Experimental results show that the proposed algorithm improves mAP50 and mAP50-95 by 4.6% and 4.2% on the AiTV dataset, and by 6.3% and 6.2% on the self-built RES dataset, respectively, while requiring fewer parameters. These gains highlight the strong practical potential of the method for real-world search-and-rescue missions.

Abstract:To address the overshoot, oscillation and slow disturbance recovery of the acoustic-pressure envelope in ultrasonic transdermal drug delivery caused by load variation, pure time delay and actuator saturation, an interval type-2 fuzzy self-tuning incremental proportional-integral-derivative (PID) control method considering load uncertainty is proposed. Based on the Butterworth-Van Dyke (BVD) equivalent circuit and envelope baseband modeling, the controlled plant is approximated as a second-order system with pure time delay. Gaussian interval type-2 fuzzy sets, Karnik-Mendel type-reduction and an anti-windup mechanism are introduced to realize bounded online tuning of PID parameters. Simulation results show that, compared with conventional PID, the proposed method reduces root mean square error (RMSE), peak overshoot, settling time and disturbance recovery time by 15.4%, 62.1%, 38.2% and 41.9%, respectively. Compared with type-1 fuzzy PID, interval type-2 fuzzy self-tuning incremental PID (IT2-FPID） further reduces peak overshoot, settling time and disturbance recovery time by 58.6%, 26.8% and 24.7%, respectively. Experimental results show that IT2-FPID outperforms the comparison methods and can achieve fast, stable and robust control for the acoustic-pressure envelope in ultrasonic drug delivery.

Abstract:Electromagnetic simulation plays a crucial role in microwave filter design, but traditional numerical methods are computationally expensive and struggle to handle the many-to-many mapping problem from design parameters to electromagnetic responses. To rapidly and accurately establish parametric models for microwave filters, this paper proposes a data-driven surrogate model based on a hybrid architecture combining conditional variational autoencoder (CVAE) and Transformer. Firstly, an S-parameter response database for a fourth-order cavity filter was constructed using electromagnetic simulation software. Subsequently, a hybrid CVAE-Transformer model was built based on this database. The model introduces a probabilistic latent space and designs an innovative composite loss function that incorporates a hierarchical decibel-domain adaptive weight mechanism, achieving a high-fidelity mapping from structural parameters to the probability distribution of S11 parameters. Experimental results show that compared to several benchmark models, the surrogate model exhibits optimal performance, with a mean squared error (MSE) as low as 0.002 324 and a correlation coefficient as high as 0.996 6. Compared to traditional electromagnetic simulation methods, the prediction speed is improved by more than three orders of magnitude. Therefore, the rapid and accurate prediction capability of the surrogate model proposed in this paper has significant theoretical guidance and practical engineering application value for the robust design, yield analysis, and optimization design of microwave filters.

Abstract:To address the challenges of significant difficulty and insufficient accuracy in locating the centerline of small-diameter buried pipelines in urban underground short and medium-distance scenarios, this paper proposes an improved Rauch-Tung Striebel (RTS) filtering and smoothing algorithm aimed at enhancing positioning precision. The method is based on a combined navigation scheme using an inertial navigation system and an odometer. It initially employs an extended Kalman filter for forward data fusion to suppress error divergence. Subsequently, an innovative improved RTS backward smoothing algorithm is proposed. By introducing a smoothed "error state", it directly corrects the positioning information and sensor parameters. Through sensor parameter compensation and secondary filtering optimization, the positioning accuracy is improved. To validate the algorithm′s effectiveness, simulation experiments were conducted. The results indicate that this method can achieve pipeline centerline mapping accuracy on the order of 10-3, fully meeting the engineering application requirements for pipeline in-line inspection positioning. This provides more reliable data support for the safe operation and maintenance of pipelines.

Abstract:To precisely address the critical issue of internal shot peening residue in slender tubes, which affects the service safety of high-end equipment, and to meet the stringent requirements for full-cycle quality assurance of core components in fields such as aerospace and precision medical devices, this study overcomes the limitations of existing nondestructive testing technologies. For the first time in the context of detecting shot peening residue inside slender tubes, this paper proposes combining eddy current testing technology with a lock-in amplifier and designs an integrated shot peening residue detection system based on field programmable gate array （FPGA） and a lock-in amplifier. Using an eddy current sensor as the probing tool, the system is inserted into the tube to capture the electromagnetic response generated by the residual shot. Based on cross-correlation theory, the lock-in amplifier enables precise extraction of weak phase signals. These signals are then quantified by a high-precision data acquisition module controlled by an FPGA, and finally, residue identification and location of the shot are achieved through phase variation characteristics. The system overcomes the application limitations of conventional ultrasonic testing, radiographic testing and single-mode eddy current testing. It can adapt to different detection scenarios through multi-frequency excitation, improving detection accuracy from the centimeter level to the millimeter level. It effectively detects residual shot of different sizes, requires no coupling agent and is unaffected by tube geometry or shot adhesion conditions. Additionally, the system offers fast detection speed, high efficiency, compact size, low cost, no radiation pollution, relaxed environmental requirements and ease of on-site deployment. A practical test platform was built to validate the detection of steel shot residue of various sizes inside aluminum-alloy slender tubes under different excitation frequencies. The results show that the detection system can effectively identify residual shots inside the tubes under different excitation frequencies. The average error in shot location detection is only 2.38 mm, with a standard deviation of 1.35 mm, and the signal-to-noise ratio is also improved compared to traditional eddy current testing methods.

Abstract:Fault diagnosis of aero-engines is of great significance to flight safety and operational efficiency. However, in practical applications, large-scale labeled fault data are difficult to obtain, leading to overfitting and insufficient generalization ability of traditional models under small-sample conditions. To address this challenge, this study introduces an advanced tabular generation model for sensor data synthesis, aiming to achieve data augmentation in small-sample scenarios. The constructed tabular generation model is extended from a tabular prior fitting network. It can quickly fit data distributions without additional training, and exhibits higher efficiency and adaptability compared with existing tabular generation methods. The prediction model is trained and evaluated using UniXDE.DTM, which is an advanced data modeling software. Experimental results show that the proposed method can effectively expand aero-engine fault samples, significantly improve diagnostic performance under small-sample conditions, and achieve optimal results in comparison with various mainstream generation models. This research provides an efficient and feasible solution for small-sample fault diagnosis, and demonstrates the application potential of large tabular models in the intelligent operation and maintenance of key industrial equipment.

Abstract:Electricity theft seriously threatens the secure and stable operation of power systems. With the continuous upgrading of metering systems, the availability of multi-source data has been significantly improved, providing favorable conditions for electricity theft detection research. However, existing methods still suffer from the limitations of single-dimensional data utilization and the difficulty of effectively capturing complex temporal dependencies. To address these challenges, this paper proposes a parallel convolutional neural network self-attention network (PCSA-Net) for electricity theft detection, which aims to exploit the complex coupling relationships among heterogeneous electricity, gas and water consumption data. In the proposed framework, multi-dimensional heterogeneous time-series data are first normalized to ensure the consistency and reliability of model inputs. PCSA-Net is composed of parallel convolutional neural networks for independent feature extraction from different energy sequences, convolutional layers for cross-channel feature fusion, and a self-attention mechanism for temporal dependency modeling and feature weighting. To alleviate the impact of class imbalance in electricity theft detection, a weighted binary cross-entropy loss function is employed during model training. Extensive experiments conducted on multiple datasets with different anomaly ratios demonstrate that PCSA-Net consistently outperforms traditional and deep learning baseline methods in terms of area under curve (AUC) and weighted F1-score (WF1), confirming its effectiveness in complex multi-source environments.

Abstract:To address the problem that software clock recovery in high-speed serial links is prone to misjudging the number of symbols between transition edges under severe jitter conditions, which further causes cumulative phase distortion in the recovered clock, an anti-distortion software clock recovery method based on phase error correction is proposed. In the software clock recovery process, the phase difference between the reference clock and the recovered clock is monitored in real time to identify and correct symbol-number decision errors between transition edges. Comparative experiments with a commercial oscilloscope show that, under conditions of high random jitter and strong duty-cycle distortion, the symbol-rate estimation error of the commercial oscilloscope can reach 0.3%, whereas the relative error of the proposed method remains below 0.005%, yielding an approximately 60-fold improvement in measurement accuracy. The results demonstrate that the proposed method can effectively suppress phase distortion in the recovered clock and improve the stability and robustness of symbol-rate estimation under severe jitter conditions.

Abstract:To address challenges in test recovery of missile-borne equipments—including uncertain recovery zones, inability to manually initiate terminal operations, unstable data transmission links and power constraints—the paper proposes and implements a self-initiating positioning and recovery system utilizing Beidou-3 radio navigation satellite service/radio determination satellite service (radio navigation satellite service (RNSS)/radio determination satellite service (RDSS)) coordination. The system adopts a centralized control architecture, employing RNSS as the primary positioning method while activating RDSS short message communication under specific recovery conditions. A five-state state machine coordination mechanism is established for "standby-awakened-flight-recovery-landing hold" operations, ensuring reliable transmission of position data and critical parameters. Additionally, a self-initiating hold circuit combining "tilt switch physical triggering+microcontroller self-hold" is designed to handle long-term storage and high-dynamic impact scenarios during launch. Motion feature analysis further distinguishes transportation disturbances from launch processes. Prototype testing demonstrates the system′s stable performance in self-initiation, state switching, RNSS positioning and RDSS data transmission. Static positioning tests show an average error of approximately 0.45 m with maximum errors below 0.85 m. Short message communication maintains an overall reception success rate of at least 86.7% in both open and obstructed environments.

Abstract:Addressing the critical challenges of weak edge representation, complex multi-scale backgrounds, and inefficient model structures in UAV-based small object detection, this paper proposes YOLO-GML—a lightweight yet powerful detector evolved from YOLOv8n. The model integrates three key innovations: LearnableDoGFilter and GaussianFilter module to enhance target contours and suppress background interference; multi-scale variable kernel convolution (MKConv） module for hierarchical receptive field fusion of local and global contexts; and a lightweight architecture achieved by pruning deep downsampling stages and optimizing the detection Head. Extensive experiments on the VisDrone2019 dataset demonstrate that YOLO-GML not only significantly reduces the parameter count to only 2.044×106 down from 3.008×106 but also achieves superior detection accuracy, with mAP50 rising from 33.633% to 42.818% and mAP50.95 from 19.628% to 25.830%. This work establishes a new leading balance between lightweight design and high precision, offering an efficient and practical solution for real-time UAV vision applications.

Abstract:In communication testing, signal sources are required to generate broadband and diverse test signals; however, the effective output bandwidth is often limited by the system clock frequency. To extend the effective output bandwidth of multiple test signals, this paper designs and implements a multimode baseband signal parallel generation system based on field-programmable gate array (FPGA）. The proposed system adopts a signal-oriented parallel generation architecture and supports the oversampled real-time generation of various baseband vector signals, including additive white Gaussian noise (AWGN), amplitude shift keying (ASK), phase shift keying (PSK), frequency shift keying (FSK), single-tone/multi-tone, chirp and comb signals. In addition, the system supports combined output of selected signal types as well as flexible configuration of multiple parameters. Experimental results demonstrate that, under a 250 MHz operating clock, the proposed system achieves an effective output bandwidth exceeding 670 MHz, while maintaining continuous modulation phase and good waveform quality, with an error vector magnitude (EVM) below 2%. The proposed system provides a feasible FPGA-based hardware implementation for the flexible generation of broadband and multi-mode baseband test signals.

Abstract:Addressing the challenging issue of error rate performance analysis and optimization in non-coherent massive single-input multiple-output (SIMO） systems, a differential amplitude phase shift keying (DAPSK） constellation design scheme based on evolutionary computation is proposed. Initially, the constellation structure is initialized according to the user modulation order and DAPSK modulation characteristics. Subsequently, an evolutionary computation framework is introduced for iterative optimization. During the evolutionary process, the initial constellation distribution is adjusted by maximizing the minimum Euclidean distance between joint constellation points, thereby reducing the computational complexity of subsequent searches. The amplitude and phase of user symbols are used as chromosome encoding parameters. Under the constraints of total power and user sub-constellation sets, the bit error rate (BER） obtained from Monte Carlo simulations is minimized as the fitness function. Candidate schemes are generated through selection, crossover and mutation operations, and the population size and evolutionary generations are dynamically adjusted to enhance convergence efficiency, ultimately obtaining the optimal constellation design scheme. Simulation results demonstrate that, compared to the baseline scheme, the proposed scheme achieves a system BER reduction of approximately 1 to 2 orders of magnitude under medium to high signal-to-noise ratio conditions and with large-scale antenna configurations. The research results demonstrate that the proposed scheme not only improve system performance but also effectively reduce detection complexity, providing a feasible solution for ultra-reliable low-latency industrial internet of things communication.

Abstract:In response to the issues of low efficiency, susceptibility to missed and false detections in traditional manual inspection of spherical shut-off valve bolts, as well as the insufficient capability of standard object detection algorithms in identifying subtle defects, this paper proposes an improved YOLOv5sbased model for bolt state detection. The model enhances feature focus on critical regions by incorporating the convolutional block attention module (CBAM） attention mechanism, optimizes detail preservation in multi-scale feature fusion through the content-aware ReAssembly of features (CARAFE） upsampling operator, and employs depthwise separable convolution (DWConv） to achieve lightweight network design. These improvements collectively enhance the identification accuracy of four bolt states: normal, loose, dislodged and offset. Experimental results on a dataset collected from hydropower plant sites show that the proposed model achieves mAP@0.5 of 93.54%, representing a 2.91% improvement over the original YOLOv5s model. Furthermore, it outperforms mainstream detection algorithms such as faster regions with convolutional neural networks (Faster R-CNN） and single shot MultiBox detector (SSD） in overall performance, demonstrating its capability to perform accurate and efficient abnormal state detection.

Abstract:To address the issues of chattering and imprecise phase compensation of the low-pass filter in traditional sliding mode observer-based sensorless control of permanent magnet synchronous motors, this paper proposes a higher-order sliding mode observer with an exponential-power switching reaching law. First, a piecewise composite function is employed to replace the sign function, effectively suppressing the inherent high-frequency chattering of the system； second, by integrating the exponential reaching law and the power reaching law, an exponential-power switching reaching law is designed to further enhance the dynamic performance and convergence efficiency of the system； furthermore, a saturation function is introduced to replace the gain of the original exponential term, mitigating the chattering caused by the exponential term near the sliding surface； finally, simulation experiments are conducted to validate the proposed method. The results show that, compared with three other typical control strategies, the errors in rotational speed and torque have been significantly reduced and the rotor position is tracked rapidly and accurately. The proposed algorithm effectively suppresses system chattering while preserving its fast response characteristics, achieving precise estimation of rotor position and speed.

Abstract:Aiming at the problem that traditional integral transform is easy to produce "ripple wave" interference in cable defect location, resulting in misjudgment and missed detection, this paper proposes a cable location compensation method based on frequency domain weighting of reflection coefficient spectrum. Firstly, the cable distribution parameter model is established, and the positioning principle based on the reflection coefficient spectrum is derived. Secondly, the attenuation constant is extracted by the resonance point and the function is fitted, and then the frequency domain weighted repair compensation function is constructed to realize the exponential gain compensation in the low frequency band and the Blackman window suppression in the high frequency band, which can effectively smooth the positioning curve and suppress the boundary oscillation. The simulation and experimental results show that the proposed method can reduce the waveform complexity in both single-defect and multi-defect scenarios, significantly eliminate the "ripple wave", make the "mutation peak" of weak defects clear and distinguishable, and has high positioning accuracy and strong anti-noise performance, which can provide a reliable basis for cable condition-based maintenance.

Abstract:Aiming at the problem of measurement accuracy of the straightness of deep hole tubes, a measurement method based on the principle of optimized by the geometric approximation searching (GAS） algorithm is proposed. Firstly, the position of the center of each cross-section is obtained step by step by the measurement device. Secondly, the least squares (LS) algorithm is used to fit the center of each cross-section to locate the reference axis. Finally, the fitted axis and straightness obtained by the LS algorithm are taken as the initial parameters, and the GAS algorithm is used to carry out the optimized iterative calculation to obtain more accurate straightness error. At the same time, the measurement software developed by laboratory virtual instrument engineering workbench (LabVIEW） can monitor the data in real time. The experimental results show that compared with the single use of the LS algorithm, the measurement accuracy of the optimized iterative combination algorithm (LS+GAS) improved by 5.24% to 19.27%, and the uncertainty of the measurement results is reduced by an average of 24.82%. The algorithm of optimized iterative combination shows high accuracy and good robustness in fitting the axis, and combined with the self-propelled device, it can significantly improve the efficiency of straightness measurement.

Abstract:Improving the prediction accuracy of wind power is of great significance to ensure the safety of large-scale wind power integration and the efficient operation of energy systems. However, the inherent non-stationarity and strong randomness of wind power series bring great challenges to prediction, and the existing models are still insufficient in the depth and breadth of feature extraction. Therefore, this paper proposes a combined prediction model based on improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN) combined with empirical wavelet transform (EWT) decomposition data, bidirectional time series convolution network (BiTCN) and bidirectional gated recurrent unit (BiGRU) prediction. The hybrid exploration selection strategy and spiral ascent factor are used to improve the global and local search performance of the dream optimization algorithm (DOA), and to search the optimal hyperparameters of ICEEMDAN and the combined model. Aiming at the noise residual problem of high-frequency components in ICEEMDAN, EWT is used to perform secondary decomposition denoising on the high-frequency components with the heaviest noise. Subsequently, the processed components are input into the combined prediction network composed of BiTCN-BiGRU to synergistically capture the multi-scale local features and long-term bidirectional dependencies of the sequence. Finally, the prediction results of each component are aggregated to obtain the final prediction value. Through the verification of the data set in this paper, the ICEEMDAN-EWT and BiTCN-BiGRU models based on improved dream optimization algorithm (IDOA） optimization are reduced to 0.73 in the average absolute percentage error, and the average absolute error and root mean square error are reduced to 2.5 and 1.46, respectively, with high prediction accuracy.

Abstract:To address the problem of low allocation efficiency caused by the strong coupling of power, bandwidth and timeslot resources in low earth orbit (LEO) satellite beam hopping systems, which further restricts the improvement of user satisfaction, this paper proposes a joint power-bandwidth allocation scheme for LEO satellite beam hopping systems based on Rainbow deep Q-network (DQN) deep reinforcement learning. First, a multi-dimensional state space is constructed by comprehensively considering features such as channel state information, queue state information and service requirements, and a joint optimization model for timeslot, power and bandwidth resources is established with the goal of maximizing user satisfaction. Second, considering the variable user demands and user states in the LEO satellite environment, the DQN is adopted for timeslot allocation learning, and the trust region interior point method is used to output power allocation to meet user requirements. Finally, to achieve and accelerate the convergence of the algorithm in the task, the Rainbow DQN algorithm is introduced in this paper.By integrating the optimization characteristics of multiple technologies, it rapidly generates the joint timeslot-power-bandwidth allocation strategy for the LEO satellite system. Simulation results show that the proposed resource allocation scheme can optimize the joint power and bandwidth configuration of the LEO satellite beam hopping system, improve user satisfaction and algorithm training stability, and effectively verify the application advantages of multi-technology fusion reinforcement learning in discrete action spaces.

Abstract:To address the non-uniform spatial distribution of pseudo point clouds, where point density is excessively high at close range but remains relatively sparse at long range, the resulting distance-dependent disparity can cause inconsistent proposal feature representations across different ranges and consequently degrade overall detection performance. To mitigate this issue, we introduce two components, namely geometric enhancement module(GEM) and adaptive region-of-interest attention fusion (ARAF) module. In GEM, a region-wise resampling strategy is adopted: voxel-level downsampling is applied to near-range regions to remove redundant points and suppress noise, while the sparse point distribution in far-range regions is preserved to maintain essential geometric structures. In ARAF, region importance is modeled at the region of interest (RoI) level by predicting adaptive weights and performing weighted fusion of features from real point clouds and pseudo point clouds. This design enables the network to dynamically emphasize far-range point cloud information according to object distance, thereby alleviating performance degradation caused by insufficient feature expressiveness in long-range scenarios. Experiments on the KITTI benchmark show that the proposed method improves the three evaluation metrics by 3.19%, 1.90% and 1.75%, respectively.

Abstract:Fatigue driving detection demands high real-time performance and reliability, yet existing object detection models suffer from large parameter counts and high computational complexity, making them difficult to deploy on in-vehicle edge devices. To tackle this problem, this paper proposes Flat-YOLO, an ultra-lightweight detection model tailored for in-vehicle environments, which reduces parameter redundancy through progressive flattened channel scaling and incorporates two complementary modules, bounding box-aware channel self-attention (BBoxCSA) and IoU-aware multi-scale fusion (IoUMSF), to enhance edge representation and multi-scale fusion. On the self-constructed driver drowsiness detection (DDD) dataset (9 751 images, 7 classes), Flat-YOLO reduces parameters by 69.4% compared to YOLOv8n while maintaining comparable mAP@0.5:0.95, with precision and recall improvements of 0.3% and 1.0%, respectively; compared to YOLOv8s, it achieves a 91.7% parameter reduction and 76.8% computational cost reduction. Class-wise results further demonstrate that for safety-critical categories, Flat-YOLO improves recall for head_low by 4.9% (0.951 vs 0.902) and for open_eyes by 2.2% (0.913 vs 0.891); for fine-grained fatigue cues, it enhances mouth′s mAP@0.5:0.95 by 0.9% (0.479 vs 0.470) and yawn′s precision by 3.3% (0.947 vs 0.914). Furthermore, on the public yawning detection dataset （YawDD） dataset, Flat-YOLO achieves an mAP@0.5:0.95 of 0.901 (compared to 0.862 for YOLOv8n), validating the cross-dataset generalization capability of the proposed method.

Abstract:Refraction and specular reflection from transparent objects severely interfere with depth-camera-based 3D measurement, resulting in holes, artifacts and noise in depth maps, which further leads to incomplete object point clouds and biased 6D grasp pose estimation. Existing depth completion methods often struggle to achieve a good trade-off between accuracy and efficiency. To address these issues, this paper proposes spectral and edge-aware depth completion network （SEDNet）, a lightweight depth completion network for robotic grasping. SEDNet first encodes robust joint representations via an RGB-D adaptive gated fusion module. It then employs an edge-aware confidence module to predict a depth confidence map, which suppresses noise in low-confidence regions during multi-scale feature fusion. A shifted-window mechanism is introduced in the downsampling stages to enhance cross-window interaction and preserve structural continuity. On the decoder side, SEDNet combines skip connections with a multi-scale spectral attention module, where differentiable wavelet transforms and dynamic spatial attention enable enhanced reconstruction of fine details and object boundaries. Experiments on the public TransCG dataset demonstrate that, compared with the mainstream baseline depth filler net （DFNet）, our method reduces root-mean-square error (RMSE）, relative error (REL） and mean absolute error (MAE） by 50.00%, 25.93% and 50.00%, respectively, while achieving an inference time of only 0.020 s per frame. These results indicate that SEDNet significantly improves depth completion accuracy while meeting the real-time requirements of robotic grasping, validating the effectiveness and practicality of the proposed method.

Abstract:Aiming at the problems of difficulty in detail recovery at low sampling rates, insufficient utilization of historical feature information and blocking artifacts in current image compressed sensing reconstruction methods based on deep unfolding networks, this paper proposes a diffusion and dual-attention alternating optimization network (DDANet) for image compressed sensing reconstruction. The network combines the diffusion model with the deep unfolding framework by designing an alternating iteration mechanism between the diffusion denoising sub-stage and dual-attention sub-stage, which achieves collaborative enhancement of data distribution priors and cross-stage feature fusion, improving the recovery quality at low sampling rates. The dual-attention sub-stage includes a cross-stage channel attention (CSCA) module and a window sparse attention (WSA) module, which enhance image reconstruction details by fusing cross-stage features and exploiting local self-similarity. To address blocking artifacts, this paper proposes a coordinate spatial self-correction (CSSC) module, which suppresses block boundary discontinuity artifacts through the collaborative action of spatial, coordinate, and self-calibration branches, ensuring global consistency of reconstructed images. Experimental results show that on the Set11 test set, when the sampling rate is 1%, the proposed method improves peak signal-to-noise ratio (PSNR） by 0.61 dB compared with diffusion message passing deep unfolding networks （DMP-DUN） which also uses diffusion model, and by 1.31 dB compared with optimization-inspired cross-attention transformer unfolding framework(OTCUF）; when the sampling rate is 10%, PSNR improves by 1.22 and 1.77 dB compared with OTCUF and unrolling fixed-point continuous network(UFC-Net） respectively, indicating that the method has significant advantages at low sampling rates.

Abstract:To address the challenges of low detection accuracy and weak small-target recognition capability for Chinese sturgeon in underwater sonar images caused by noise interference, this study proposes an improved YOLOv8n-based sonar detection model. The methodology integrates three key innovations: a hierarchical attention fusion block module is designed to achieve adaptive fusion of local and global features through a hierarchical attention mechanism; a pinwheel-shaped convolution module is introduced, employing a multi-branch convolution fusion strategy with a Gaussian-distributed receptive field to enhance differential feature extraction for small targets; a lightweight shared convolutional separated BN (batch normalization) detection head is developed to balance detection accuracy and inference efficiency. Experiments are conducted on a self-constructed Chinese sturgeon sonar dataset, partitioned into training, validation, and test sets at an 8∶1∶1 ratio. Experimental results demonstrate that compared to the baseline YOLOv8n model, the proposed approach improves mAP@0.5 by 1.0%, precision by 2.6%, and recall by 0.3%, while reducing parameters by 0.58×106 and floating pointoperations by 0.7 GFLOPs. Ablation studies confirm that the synergistic integration of the three modules significantly enhances model performance. Comparative experiments show that the proposed model outperforms mainstream object detection models in detection accuracy under stable parameter complexity. The improved model achieves notable enhancements in both detection precision and efficiency for Chinese sturgeon in complex underwater noise environments, providing an effective technical solution for underwater sonar behavior monitoring. This work holds significant application value for Chinese sturgeon conservation, wild training in artificial breeding, and population restoration efforts.

Abstract:To address the problem that coal and gangue have similar appearances under complex working conditions, and that illumination fluctuation, motion blur, and occlusion easily lead to low detection accuracy as well as high false detection and missed detection rates, a coal gangue object detection method based on ECA-DEIM is proposed. In the backbone network, a lightweight multi-scale attention module is introduced to enhance the capture capability for multi-scale targets. A coal-gangue spatial refinement and selective correction module is constructed to improve the spatial representation capability of similar targets. An adaptive dynamic normalization module is designed to enhance the generalization ability of the model. Experimental results show that, on a self-built dataset, the mAP@［0.5:0.95］ and mAP@0.5 reach 52.3% and 82.5%, respectively, which are 2.2% and 2.4% higher than those of DEIM, while the number of parameters is reduced by 19.9%. The missed detection rate is reduced to 3.95%, the false detection rate is decreased by 4.3%, and the sorting accuracy reaches 93.1%. The results indicate that the proposed method has good detection accuracy, robustness, and engineering applicability under complex working conditions.

Abstract:To address the challenges of limited feature extraction capability, dense distribution of small objects, and drastic target scale variations that lead to low detection accuracy in unmanned aerial vehicle (UAV) aerial imagery, this paper proposes an improved UAV object detection algorithm based on real-time detection transformer (RT-DETR). Firstly, in the backbone network, a C2f-EncoderBlock feature enhancement module is designed to restructure the network architecture. By introducing a frequency-spatial domain collaborative perception mechanism and utilizing fast Fourier transform （FFT） to capture global illumination and texture information in the frequency domain, this module effectively resolves the difficulty of feature extraction in low-resolution and blurred images while significantly reducing the number of model parameters. In the detection head, a multi-scale small object enhancement pyramid structure is designed to effectively address the issues of missed detections for small objects and coarse feature fusion in the original model through lossless downsampling and multi-dimensional feature interaction. Finally, a focaler minimum point distance intersection over union (Focaler-MPDIoU) loss function incorporating a hard-sample focusing mechanism is designed to replace the original generalized intersection over union (GIoU). This function not only improves the geometric localization accuracy of dense targets by minimizing corner distances but also enhances the model′s ability to mine hard samples, such as tiny or occluded objects, through a dynamic re-weighting strategy, thereby further improving detection accuracy. Experimental results on the VisDrone2019 dataset demonstrate that the improved algorithm achieves a 3.9% increase in mAP50 compared to the baseline model, while reducing the parameter count by 24.75%, effectively enhancing UAV object detection performance.