Abstract:To address the insufficient spatio-temporal feature perception caused by neglecting historical traffic information in existing methods,this study proposes an intersection signal control method integrating deep reinforcement learning with spatio-temporal feature modeling. The approach employs a hybrid D3QN-LSTM network architecture,which encodes multi-period traffic information into high-dimensional matrices through discrete traffic state representation. A convolutional neural network extracts spatial features,while a long short-term memory network captures temporal dependencies. A reward-feedback-driven dynamic exploration mechanism is further designed to optimize policy training. Experiments conducted on the SUMO simulation platform demonstrate that during morning peak traffic,the proposed method reduces average queue length by 49.95%,35.04% and 16.72%,and decreases cumulative waiting time by 63.03%,35.55% and 20.15% compared to fixed-timing control,conventional reinforcement learning methods and D3QN,respectively,validating the superiority of spatio-temporal feature modeling and dynamic exploration strategies. To assess algorithmic robustness,off-peak traffic flow experiments further confirm that the proposed method maintains significant advantages in both average queue length and cumulative waiting time metrics,demonstrating strong adaptability and generalizability across varying traffic load conditions.

Abstract:The temperature measurement error caused by solar radiation, that is, the solar radiation error, can be as high as 1 K. To improve the temperature measurement accuracy, and to address the problem of high power consumption of traditional forced ventilation temperature measurement device, a temperature sensor device for brushless DC (BLDC) fan ventilation is designed, which combines natural ventilation and forced ventilation functions, effectively reducing power consumption. The computational fluid dynamics (CFD) method is utilized to conduct multi-physical field fluid-structure coupling simulation of the sensor and quantify the radiation error under different conditions, and the functional mapping relationship between wind speed and solar radiation intensity and the suction pressure of the wind turbine is established by minimizing the radiation errors, and formulate the wind turbine control strategy. The genetic algorithm-optimized BP neural network (GA-BP) algorithm optimized by genetic algorithm was compared and selected to train and fit the simulation data set, thereby constructing the radiation error correction equation. Finally, through the field comparison experiment with the 076B temperature sensor, it is shown that the radiation error of the designed temperature sensor after algorithm correction can be controlled within 0.05 K, the mean absolute error was 0.039 K, and the root mean square error was 0.045 K.

Abstract:Considering the problem of high failure rate and difficult maintenance of imported traveling wave tube amplifier, an indoor Ku-band 350 W solidstate power amplifier was developed. A novel 16-way power dividing/combining network was proposed, which was based on novel waveguide magic T, waveguide E-plane T-junction, coplanar magic T and half-height waveguide-to-microstrip probe transition. A Ku-band 450 W power amplifier module was achieved based on 16 pieces of 35 W gallium nitride power amplifier chips and the 16-way power dividing/combining network. Then, an indoor Ku-band 350 W solid-state power amplifier was successfully developed. The measured results show that the gain is greater than 73 dB, the output power is greater than 400 W between 13.75 GHz and 14.5 GHz. After calculation, the overall efficiency is 23.65% at rated power output, which is on par with international well-known companies. This power amplifier has excellent specifications and can replace imported TWTA products completely.

Abstract:It is a desired solution to integrate photovoltaics (PV) and battery into the high-voltage dc bus using high-gain three-port converters in renewable energy systems. Aiming at the limitations of conventional three-port converters—including restricted voltage gain, difficulty in achieving soft switching, and high voltage stress on semiconductor devices—this paper proposes a novel ultra-high voltage gain three-port converter topology based on a three-winding coupled inductor. Only one magnetic core is used so that the power density of the converter is effectively improved. The lower voltage stress and soft switching performance of semiconductor devices enable specifications with lower conduction losses to be selected, which can reduce system losses and improve efficiency. Based on the port power relationship, the proposed converter can achieve smooth switching between different operating modes. The topology and operating principles of the converter are analyzed in detail, and then the port voltage relationship, voltage/current stress, and control methods are analyzed in detail to guide parameter design. Finally, both simulation models and experimental prototypes were developed with PV input voltages ranging from 20 V to 40 V, a battery voltage of 48 V, an output voltage of 400 V, and a rated power of 400 W, validating the effectiveness of the proposed high-gain three-port converter and its control strategy.

Abstract:The temperature characteristic has significant influence on the dynamic performance of magnetorheological damper, so it is of great significance to study the theoretical model of temperature rise, to analyze and improve magnetorheological damper temperature characteristic. Based on the energy balance relationship of magnetorheological damper, the temperature rise theory and heat transfer mechanism of magnetorheological damper under different motion amplitudes excitation are revealed, and the temperature rise theoretical models under small and large sinusoidal harmonic motions are established.The temperature rise rule of damper is analyzed by finite element simulation. The simulation results show that the temperature of magnetorheological fluid in the damper cavity increases to 3.2℃ with small sinusoidal harmonic motion amplitude, and the temperature of magnetorheological fluid at different positions in the damper cavity is large. The temperature of magnetorheological fluid increased by 20.8℃ during the large motion amplitude, and the temperature of magnetorheological fluid at different positions in the cavity is almost equal, the theoretical model of temperature rise of magnetorheological damper is verified. It is verified by the temperature rise characteristic test, and the temperature rise test curve is consistent with the simulation and theoretical calculation results. The theoretical model has a large error in predicting the temperature rise of magnetorheological fluid under small motion amplitude, while the predicted value is more accurate under large motion amplitude. The temperature rise theoretical model not only effectively predicts the internal temperature of the damper, but also provides a theoretical basis for the structural design and engineering application of magnetorheological damper.

Abstract:To address the issues of target detection accuracy degradation and small target miss detection caused by speckle noise interference, low signal-to-noise ratio, and multi-scale scattering characteristics of targets in Synthetic Aperture Radar images, this paper proposes a lightweight detection model named XMNet, which balances feature representation capability and real-time performance. XMNet incorporates an improved single-Head vision Transformer into the backbone network to strengthen contextual semantic correlations through global attention mechanisms. A cross-layer multi-path aggregation network is designed as the neck structure, integrating dynamic upsampling and a parallel multi-scale convolution module to optimize multi-scale feature representation. An additional high-resolution detection layer is introduced to leverage shallow high-resolution features, enhancing detail capture capability for small targets. Experiments on the MSAR-1.0 dataset demonstrate that XMNet achieves a mean average precision of 90.4% across all categories, representing an increase of 8.7% over the baseline model. Detection accuracy for small aircraft targets significantly improves by 20.1%, with only a 2-million parameter increase while achieving an inference speed of 185 FPS. When compared against nine advanced methods including FCOS and CenterNet, XMNet ranks first in comprehensive metrics balancing detection accuracy and computational efficiency. Through the design of cross-layer attention mechanisms and multi-scale feature fusion, XMNet effectively resolves the challenge of balancing feature preservation for multi-scale targets and real-time processing in SAR imagery. Its lightweight and high detection accuracy provide a viable engineering-ready solution for real-time remote sensing monitoring across various SAR platforms, demonstrating significant advantages particularly in complex scenes with dense small targets.

Abstract:To address the current limitations in data-driven remaining useful life (RUL) prediction methods for turbofan engines, which suffer from low data utilisation and constrained prediction accuracy due to inadequate exploitation of data feature information, a novel multi-scale RUL prediction model for engines is proposed. This model is termed the distributed spatio-temporal convolutional network (DSCN). The proposed method first captures linear and non-linear relatonships in engine data by calculating Pearson correlation coefficients and maximum information coefficients, thereby obtaining trend features for both stationary and non-stationary time series. Secondly, it employs a multi-scale residual fusion module to enrich data features. Building upon temporal convolutional network (TCN), it incorporates residual channel attention module (Res-CAM) and multi-head attention module (MHAM) to enhance the model′s ability to capture critical information, dynamically adjusting the weights of the data. The proposed method was experimentally validated on the FD001 and FD003 datasets within the C-MAPSS collection, yielding RMSE and Score values of 11.30 and 218.08; 12.04 and 227.65 respectively. Results indicate that this approach reduces the Score by 4.67% and 11.5% compared to the current state-of-the-art method.

Abstract:To address the limitations of existing CNN-based generative steganography in poor image quality and weak resistance to steganalysis, this paper proposes SSEU-Net, an improved U-Net-based steganographic architecture incorporating selective state space model, aiming to achieve high-quality image generation and secure steganography. The core contributions include: first,designing Res-SS2D module that performs quad-directional global spatial modeling on input images while maintaining linear computational complexity, thereby enhancing the visual quality of stego images; next, proposing a high-frequency feature enhancement strategy based on the observation that subtle perturbations in high-frequency regions minimally affect statistical characteristics. This strategy extracts and integrates edge features of carrier images into the encoder to guide secret information embedding into high-frequency regions, thereby reducing detectability by steganalysis; finally developing a multi-objective loss function combining PSNR and MS-SSIM for generation quality optimization, alongside introducing an L1 norm loss on low-frequency components to enforce consistency between cover and stego images in low-frequency regions, ensuring secret information is predominantly embedded in high-frequency components. Experiments demonstrate that SSEU-Net outperforms existing methods on COCO and ImageNet datasets. On ImageNet, the generated stego images achieve an average PSNR of 40.588 dB, with extracted secret images attain an average PSNR of 41.863 dB, while exhibiting strong resistance to common steganalysis.

Abstract:After long-term use of Vicat softening temperature testers, problems such as temperature indication deviation and inaccurate heating rate often occur. To address this, a high-precision calibration device based on the principle of dynamic testing has been developed. This device is equipped with an auto-triggered image acquisition and temperature recording device, which can real-time capture the deformation critical point and accurately record the temperature at the moment when the sample undergoes a 1 mm deformation or is pierced by the pressure needle. Its working process includes steps such as sample placement and parameter setting, with smooth connection between each step. In practical tests, the device shows obvious advantages: at a heating rate of 12℃/6 min, the actual heating rate ranges from (11.9~12.1)℃/6 min, and the temperature indication error is (-0.1~0.2)℃; at a heating rate of 5℃/6min, the actual heating rate is in the range of (4.9~5.1)℃/6 min, and the temperature indication error is (-0.1~0.2)℃, all meeting the technical requirements. Through analysis, the standard uncertainty of temperature indication error is 0.067℃, and the standard uncertainty of heating rate error is 0.091℃/h. The innovation of this research lies in the proposal of a dynamic heating calibration method, optimization of dynamic acquisition strategy, and integration of intelligent calibration algorithms with automated processes, which can support synchronous calibration of multiple temperature points and multiple sample stations. This device effectively solves the problems existing in current calibration methods, realizes synchronous dynamic and accurate measurement of temperature and heating rate, provides a reliable guarantee for material thermal performance testing, and plays an important role in fields such as material research and development and quality control.

Abstract:To address the issues of trailing signals generated by ultrasonic transducers under excitation pulses and reverberation interference caused by multiple reflections and scattering of acoustic waves on the tank wall, which lead to a large blind zone in ultrasonic ranging, this paper proposes the use of Linear Frequency Modulated (LFM) waves with anti-reverberation capability as the transmit signal. Furthermore, aiming at the problems of image spectrum generation resulting in target detection ambiguity and high computational load when traditional receivers directly acquire real signals, quadrature demodulation technology is adopted at the receiver end. This approach not only obtains complex signals with strong anti-interference capability but also reduces system costs. By analyzing the reverberation model of the ultrasonic level meter and comparing the ambiguity function and Q-function of CW and LFM waves through simulation, this paper concludes that LFM waves possess superior target resolution capability and better anti-reverberation performance when the target is stationary. Experiments were conducted using LFM waves as the transmit signal. At the receiver end, complex signals were obtained via quadrature demodulation, followed by matched filtering. Experimental results demonstrate that the maximum absolute measurement error of this method is less than 4 mm, and the blind zone can be reduced to 8 cm, indicating high practical value and engineering significance.

Abstract:Existing BCI neurofeedback techniques often struggle to balance temporal and spatial resolution. Among mainstream neurofeedback methods, EEG offers millisecond-level temporal resolution but lacks precise spatial localization, whereas fMRI provides high spatial resolution but is constrained by second-level temporal delays. This trade-off in spatiotemporal resolution limits the clinical applicability of neurofeedback. To address this issue, this study proposes a hybrid wavelet neural network to model the complex nonlinear mapping between EEG signals and fMRI regional activity. The model employs parallel wavelet convolutional layers and one-dimensional convolutional layers to extract multi-resolution frequency-domain features and local time-domain features from EEG signals, respectively. A channel cross-attention mechanism is further introduced to capture nonlinear interactions between features, while a LSTM network models long-range temporal dependencies. Experimental results demonstrate that the proposed approach achieves high-precision prediction of fMRI regional dynamics across two independent datasets, significantly outperforming traditional linear models. This framework not only extends the modeling capacity of current neurofeedback “EFP” techniques but also provides a new pathway for developing neurofeedback and BCI systems with both high temporal and spatial resolution.

Abstract:Accurate lesion segmentation is crucial for early diagnosis and subsequent treatment of dermatological diseases. Existing neural networks often employ increasingly deep and complex architectures to achieve high segmentation accuracy; however, large parameter counts and high computational costs limit practical deployment. To address these challenges, a lightweight multi-scale channel interaction segmentation network (LMSCI-Net) is proposed. For each input image, a lightweight multi-scale encoding module based on channel separation and convolutional decomposition is designed, augmented by a local-global channel attention mechanism to ensure robust feature extraction while maintaining an efficient encoder. A multi-scale channel interaction enhancement module is then introduced to integrate multi-stage outputs and refine skip connections, providing the decoder with rich and precise detail information. Finally, an adaptive fusion decoding module is developed to progressively restore fine-grained details and produce accurate segmentation masks. The network is trained under a deep supervision regime and evaluated on three public skin-lesion segmentation datasets (ISIC2017, ISIC2018 and ISIC2016) as well as the PH2 dermoscopic image database. Experimental results demonstrate that, compared with the U-Net baseline, LMSCI-Net reduces parameter count and computational complexity by 99.38% and 98.78%, respectively, while maintaining high segmentation accuracy and strong generalization, thus validating its effectiveness and lightweight design.

Abstract:UAV aerial photography is one of the mainstream object detection technologies, and this task faces problems such as small target objects, large scale changes, and complex background interference. How to improve the detection accuracy with limited computing resources is an important challenge. In order to solve the above problems, a lightweight UAV aerial target detection method was proposed. Firstly, a hierarchical dependence-aware pruning algorithm was designed to reduce the redundancy of the model. In addition, the resolution of the detection head is increased to 160×160 to enhance the detection ability of small targets, the standard convolution blocks of the network are replaced by GhostConv to reduce the computational redundancy, and the C3 module in the Neck network is redesigned by introducing the compact architecture StarNet to reduce the complexity of the feature fusion process and enhance the feature expression ability. Finally, the attention mechanism is introduced in the backbone layer to improve the feature extraction ability of the model. The experimental results show that in the VisDrone2019 dataset, the mAP_0.5 of the model is increased by 1.8%. At the same time, compared with the original model, the number of parameters is reduced by 50.4%, and the amount of computation is reduced by 35.44%. In summary, the model satisfies the requirements of the UAV platform for accuracy and lightweight in small target detection tasks.

Abstract:Aiming at the multi-point path planning problem for electromagnetic interference testing, a path planning method based on the combination of improved A-star algorithm and grey wolf optimization algorithm is proposed. First, the traditional A-start algorithm is improved by modifying the heuristic function and introducing a redundant point deletion strategy, thereby reducing path length and algorithm runtime. Then, the test path planning problem is transformed into a classic traveling salesman problem and solved using the improved gray wolf optimization algorithm to obtain the optimal test path. Experimental results demonstrate that compared to traditional methods, the improved approach achieves an average reduction of 4.73% in total path planning distance, 30.42% in average number of turns, 34.74% in average total turning angle, and 39.47% in average computation time. This effectively enhances testing efficiency and safety, providing a reliable solution for electromagnetic interference multi-target point testing tasks.

Abstract:To address the issues of existing object detection models being prone to interference and resulting in insufficient accuracy when detecting barcodes in complex environments, as well as the high model complexity making deployment on low-computing-power mobile devices challenging, this study proposes a lightweight, high-precision detection algorithm called DOLN-YOLO based on YOLOv8. First, the DW-HGNetV2 architecture, reconstructed using deeply separable convolutions, is introduced as the backbone network, which enhances multiscale feature extraction capabilities while significantly reducing computational complexity. Second, the OD-C3Ghost module is constructed to replace the C2f module, enhancing dynamic perception capabilities for complex barcode deformations and further eliminating computational redundancy. Third, a lightweight shared detail enhancement detection head is designed, utilizing the gradient-strength dual-channel coordination mechanism of DEConv to enhance the model′s feature generalization capabilities, and adopts a heterogeneous convolution sharing strategy to reduce resource consumption; finally, a composite loss function NWD-PIoUV2 is proposed, combining normalized Wasserstein distance with dynamic focus PIoUV2 loss, to mitigate the optimization challenge of minor localization deviations and accelerate convergence speed. Experimental results demonstrate that, compared to the baseline model, DOLN-YOLO achieves a 0.92% improvement in mAP@0.5 and a 4.57% increase in mAP@0.5:0.95, while reducing parameters and computational costs by 58.8% and 48.6% respectively. This validates the algorithm′s superiority in detecting barcodes under complex environments. DOLN-YOLO provides a solution featuring both robust detection capability and efficient mobile deployment for logistics, healthcare, retail, and other application scenarios.

Abstract:To address the issues of imprecise target localization, missed detections, and false alarms caused by significant differences in target scales, diverse categories, and uneven target distribution in remote sensing images, this paper proposes an improved algorithm based on YOLOv8n, named MGD-YOLO. Firstly, the multi-scale edge-gaussian attention module （MEGA） is introduced. By integrating Gaussian smoothing, the Scharr edge operator, and a channel attention mechanism, MEGA effectively suppresses noise and enhances the feature representation of target contours in complex backgrounds. Secondly, the MDPConv structure is designed, which combines a dynamic weighted fusion mechanism with depthwise separable convolutions to overcome the fixed receptive field problem of traditional convolutions and improve the model′s ability to detect targets of varying scales. Lastly, the DLGA structure is introduced in the detection head. By dynamically allocating weights to multiple attention branches and utilizing an MLP fusion strategy, DLGA significantly improves the integration of local and global features, thereby boosting detection performance. Experimental results demonstrate that MGD-YOLO achieves a 1.6%, 2.7% and 1% increase in mAP@0.5 on the DIOR, DOTA and NWPU VHR-10 datasets, respectively, compared to YOLOv8n, thus validating its effectiveness for remote sensing image target detection tasks.

Abstract:To address the widespread instability and volatility in stock price forecasting, as well as the difficulty of parameter optimization in the variational mode decomposition (VMD) algorithm, this paper proposes a two-stage combined prediction framework, CRIME-SE-VMD-VIT2M. In the first stage, the Chebyshev chaos map and lens imaging population selection strategy are introduced on the basis of the original frost ice optimization algorithm. Using SE as the fitness function, an improved CRIME-SE-VMD optimization model is constructed to enhance the global search capability and decomposition quality of parameter optimization. In the second stage, key technical indicators are selected through PCC and fused with the IMFs obtained from VMD decomposition to form a multi-dimensional feature set. Based on this, combined with the optimization results of the first stage, a VIT2M parallel dual-channel prediction model is designed and implemented to deeply extract and model multi-scale stock feature information. Experimental results show that the fitness value of CRIME-SE-VMD on four stock datasets is 0.000 318 9~0.000 703 lower than that of the comparison algorithm, demonstrating better decomposition performance. At the same time, the prediction performance of the VIT2M model on the same datasets is better than that of the comparison model, verifying its effectiveness in improving the accuracy of stock price prediction.

Abstract:To address the issue of missing dynamic interactions between units caused by existing intent recognition methods failing to account for formation spatial characteristics, this paper introduces a spatiotemporal coupling mechanism and proposes a formation intent recognition method that integrates dynamic graph attention mechanisms with spatiotemporal modeling. First, a global interaction backbone network is established based on target attributes within the formation, combined with a Top-K nearest neighbor strategy to dynamically generate an adjacency matrix. This transforms the dynamically evolving formation state into a structured temporal graph. Second, a Graph Attention Network (GAT) enhanced with a Co-evolution Aware Pooling mechanism is employed to adaptively learn differentiated interaction weights between different targets, enabling rapid capture of spatial coordination features in the dynamic formation temporal graph. Finally, a Bidirectional Gated Recurrent Unit (BiGRU) augmented with a MultiScale Attention mechanism is introduced to perform temporal analysis on the node state sequences extracted by GAT, which contain spatial coordination information. This establishes an intent recognition model that deeply integrates spatiotemporal features (STGAT-BiGRU). Simulation results show that, compared to existing methods, the proposed approach achieves average improvements of 8.78% and 8.9% in accuracy and F1 score, respectively, demonstrating its effectiveness and providing technical support for mastering situation evolution and gaining decision-making initiative.

Abstract:To address the challenges of low manual efficiency and surface damage risks in contact-based measurement for precision optical slits and pinhole lenses, this paper proposes an interpolation-Zernike collaborative subpixel detection method. By enhancing edge resolution through bicubic interpolation, reducing discrete sampling errors via reconstructed orthogonal basis templates, and correcting subpixel offsets with an asymmetric Gaussian model, the method improves antiinterference capabilities through dynamic thresholding and small connected-domain denoising. Simulation and experimental results demonstrate that the improved algorithm achieves a maximum detection error of 0.098 7 pixel (1.401 5 μm) for slit width and stabilizes pinhole diameter errors within 0.12 pixel (1.704 μm), representing a 62.3% accuracy improvement over traditional Zernike moment methods. Under pixel-aligned conditions, the method achieves nanoscale resolution of 0.000 2 pixel (2.84 nm), surpassing conventional micron-level limitations. The algorithm exhibits a linear positive correlation between accuracy and camera resolution, meets industrial detection standards within 3 μm under the experimental conditions, and demonstrates potential for nanometer-scale applications. This work provides an innovative solution for high-efficiency, non-destructive inspection of optical components.

Abstract:Sparse angle CT is an effective method to reduce X-ray radiation dose in clinical CT imaging. However, due to the incomplete projection caused by sparse sampling, the image reconstruction contains obvious fringe artifacts. In order to solve this problem, this paper proposes a sparse angle CT image reconstruction network based on iterative optimization deployment, IADR-Net, which adopts a unique dual-channel parallel architecture design, and includes two core components: Iterative reconstruction sub-network and global-local attention network (GLONA) detail recovery sub-network. Among them, the iterative reconstruction sub-network is based on the framework of fast iterative soft threshold algorithm, and realizes projection-to-image reconstruction through learnable nonlinear transformation and adaptive thresholding. The GLONA sub-network adopts a double-branch structure with parallel local and global features, and effectively maintains the image details through the self-adjusting fusion module. The two sub-networks work together to focus on artifact removal based on iterative expansion and detail enhancement based on attention mechanism, respectively, and finally output high-quality CT images through feature fusion. Experimental results on the Mayo dataset show that the proposed method has better performance than several representative algorithms in terms of artifact suppression and structure preservation, and provides an effective solution for clinical sparse angle CT imaging.

Abstract:Dense target distribution, complex backgrounds, and a large number of small objects often lead to suboptimal detection performance in remote sensing image object detection. To address these challenges, this paper proposes RSD-DETR, a remote sensing object detection algorithm based on RT-DETR. First, a lightweight multi-scale feature extraction module, Faster-CGLU, is designed by integrating a gating mechanism with partial convolution, which optimizes the aggregation of local and global feature information while reducing computational redundancy. Second, a CGA-AIFI module is constructed using cascaded group attention (CGA), which focuses on critical feature regions while suppressing irrelevant background information, thereby enhancing the interaction between the model and object features. Finally, a cross-scale dynamic feature fusion module (CS-DFFM) is designed, which performs spatial alignment and dynamic fusion of multi-scale feature maps through the dynamic scale-sequence feature fusion (DySSFF) module and the triple feature encoder (TFE) module. This effectively mitigates the loss of small object features caused by upsampling and downsampling, and enhances the network′s multi-scale feature fusion capability. Experimental results show that on the SIMD and DOTA-v1.0 datasets, the proposed algorithm reduces the number of parameters by 22.11% compared with the baseline model, and the mean average precision (mAP0.5) reaches 79.9% and 86.8% respectively, which are 2.5% and 1.7% higher than those of the baseline model. The real-time performance of the model is also improved. The detection effect is better than other classic models, and it has excellent performance.

Abstract:Aiming at the challenges of remote sensing image registration such as feature extraction difficulties caused by complex environment and registration accuracy limitations caused by multi-scale geometric deformation, this paper proposes a remote sensing image registration model that integrates dual-domain features and cross-dimensional gated attention. Firstly, the multi-scale Fourier module is designed in the feature extraction stage to improve the StarNet network structure to enhance the feature extraction capability of the model by fusing the multi-scale spatial features with the frequency domain features; then, the cross-dimensional gated attention is designed so that the model can efficiently capture the contextual information in the image without sacrificing the global sensing field; secondly, the feature matching stage bidirectional parameters are obtained by applying bidirectional matching based on partial assignment matrix, and finally, the registration is completed by affine transformation. In the experiments using the aerial image dataset, the results show that when the correctly estimated keypoint scale factor is set to 0.01, 0.03 and 0.05, the registration accuracy reaches 42.8%, 85.7% and 96.9%, respectively, and the average registration time is 0.87 s, which significantly improves the accuracy and speed of remote sensing image registration.

Abstract:Accurate monitoring of activated sludge microorganisms is critical for maintaining the stable operation of wastewater treatment systems. However, due to their semi-transparent morphology and high similarity to the surrounding environment, these microorganisms exhibit camouflaged characteristics, rendering traditional detection methods ineffective. To address the camouflage characteristics of activated sludge microorganisms, the diversity of object scales, and the ambiguity of boundaries in complex contexts, this paper proposes a camouflaged object detection method based on multi-scale awareness and edge enhancement. The proposed method employs a multi-scale feature aware module to extract rich contextual information through parallel processing and progressive expansion of the receptive field, thereby enhancing multi-scale feature representation. An edge-aware enhancement module is introduced to fuse low-level edge details with high-level semantic information for more accurate edge feature extraction. These edge features are then integrated with the multi-scale features through an attention-guided feature module, enabling the network to focus on the positional information of edges. Finally, a context aggregation module is used to progressively aggregate multi-level features in a top-down manner, further refining the prediction and generating the final output. On the benchmark camouflaged object detection dataset and the self-constructed activated sludge microorganism camouflage dataset, the proposed method achieves improvements of 2.2%, 4.1%, and 2.1%, and 1.2%, 2.2%, and 0.6% in terms of the evaluation metrics S-measure, weighted F-measure, and E-measure, respectively. Experimental results demonstrate that the proposed method achieves superior performance over other models across all datasets.

Abstract:Due to light scattering in water, underwater images commonly suffer from quality degradation. To address this issue, this paper proposes an enhancement model for turbid underwater polarized images based on an enhanced LU2Net network, validated using a self-constructed dataset. Initially, the acquired color polarization images are converted to grayscale. Complete linear polarization information is obtained by fusing the three polarization components at 0°, 45° and 90°. The degraded underwater polarized images are subsequently enhanced using the proposed enhanced LU2Net network model. Finally, enhanced images possessing richer detail features are acquired. Experimental results demonstrate that the proposed method outperforms comparative underwater image enhancement techniques including FUnIE-GAN and MLLE, in terms of both subjective and objective evaluations, as well as in the outcomes of feature point detection and Canny edge detection. Crucially, during feature point detection employing four distinct methods including ORB and AKAZE, the proposed approach consistently extracted a greater number of feature points.The proposed method achieves a 3.35% reduction in LPIPS compared to the best-performing existing method used for comparison. Furthermore, it increases the UCIQE score by 1.16% and decreases the NIQE score by 7.59% compared to the algorithm prior to enhancement. The proposed method successfully extracts clearer image edges, textures, and other fine details in turbid water environments under natural lighting conditions, thereby enhancing imaging quality in such challenging scenarios.

Abstract:To address the challenge of balancing accuracy and real-time performance in front-end target recognition and localization for drones in complex battlefield environments with limited onboard resources, a front-end target recognition and localization method for drone operations in complex battlefield environments was developed: Using a ″backbone-neck-head″ as the basic network architecture, a non-local attention expansion module, a global multi-scale decoupled network, and a lightweight bottleneck module were introduced. Focal Loss and DIoU Loss were employed as the combined loss functions to achieve feature modeling and multi-scale detection enhancement, thereby improving the ability to capture features and enhancing accuracy; based on dependency graph-structured pruning and channel-wise knowledge distillation, a collaborative lightweight strategy was proposed, effectively reducing model complexity and improving embedded deployability. Experiments show that this method improved mAP@0.5, mAP@0.75, and mAP@0.5:0.95 by 6.0%, 7.2%, and 5.9% respectively, while reducing model parameters and GFLOPs to 17.1% and 12.0%, with precision loss controlled within 4.1%. Finally, deployment validation on embedded hardware demonstrated a frame rate of 34 fps, effectively meeting the accuracy and real-time requirements for front-end target recognition and localization during drone operations.