• Volume 58,Issue 9,2026 Table of Contents
    Select All
    Display Type: |
    • Design of variable fractional delay filter in broadband multi-beam forming

      2026, 58(9):1-8. DOI: 10.11918/202412039

      Abstract (491) HTML (55) PDF 8.09 M (141) Comment (0) Favorites

      Abstract:To reduce the performance loss caused by fixed-point quantization of variable fractional delay filter coefficients using truncation methods and optimize the resource consumption of fractional delay filter engineering applications, this paper proposes a Farrow filter design method that combines the Minimax method and mixed integer linear programming(MILP). And, the multiplexing method of Farrow filters is introduced. Firstly, based on the Farrow filter structure and coefficient symmetry, the filter coefficient solution is characterized as an optimization problem with amplitude-frequency response peak error constraints and fractional delay error constraints and integer constraints are established on the Farrow filter coefficients. Secondly, based on the established constraints, the filter coefficients are solved. Then, resource consumption is analyzed based on the introduced Farrow filter reuse method. Finally, a 16-channel, 4-beam beamforming system experimental platform is built on the Xilinx VCU118 development board to validate the filter design method and multiplexing method proposed in this paper. The experimental results show that, under the same quantization bit width conditions, compared with traditional truncation quantization methods, the proposed method reduces the normalized root-mean-squared and group delay error of fractional delay filters. The Farrow filter bank has significantly optimized resource consumption after reuse. The digital beam pattern formed by the experimental platform meets expectations without any directional deviation. The design method and multiplexing method proposed in this paper can provide reference for the engineering application of Farrow filter in broadband multi-beamforming system.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
      • 11+1
      • 12+1
    • Low-complexity single-snapshot DOA estimation algorithm for one-dimensional linear arrays

      2026, 58(9):9-16. DOI: 10.11918/202412051

      Abstract (205) HTML (26) PDF 3.23 M (57) Comment (0) Favorites

      Abstract:To address the limitations of existing single-snapshot direction of arrival (DOA) estimation algorithms, namely the high computational complexity in real-time applications and the restriction to uniform linear arrays, this paper proposed a low-complexity single-snapshot DOA estimation algorithm applicable to arbitrary linear arrays. The algorithm adopted a two-stage strategy of initial estimation+optimization iteration. In the initial estimation stage, an extended adaptive diagonal loading technique was employed to mitigate the divergence problem of noise eigenvalues caused by insufficient snapshots, and approximate algorithms were introduced to replace traditional matrix decompositions to reduce computational complexity. At the same time, an estimation-subtraction strategy was incorporated to ensure that the loading matrix focused on a single source per estimation, thereby improving angular resolution and matching the loading level to enhance estimation accuracy. In the optimization stage, the high-accuracy characteristic of the initial estimation results was leveraged to perform fast iterations within a smaller search range, in order to further improve estimation performance and reduce computational cost. Simulation results demonstrate that under the conditions of uniform linear arrays and non-uniform linear arrays, the proposed algorithm outperforms the five benchmark algorithms in terms of DOA estimation accuracy and angular resolution; its computational complexity is the lowest, and the average computation time is reduced by 69.60% and 76.94%, respectively. While ensuring high-accuracy DOA estimation performance, the proposed algorithm significantly reduces the computational complexity, and can provide effective technical support for target localization and parameter estimation in highly real-time systems such as radar, communications, and underwater acoustic direction finding, holds high engineering application value and promotion significance.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
    • Incremental intrusion detection mechanism considering cost-sensitive

      2026, 58(9):17-26. DOI: 10.11918/202412064

      Abstract (173) HTML (28) PDF 6.68 M (65) Comment (0) Favorites

      Abstract:Aiming at the issues of data imbalance and forgetting imbalance during training, which makes incremental intrusion detection system prone to missed detection of the specific attack classes with high-cost misclassification, an incremental intrusion detection mechanism considering cost-sensitive (IIDM-CCS) is proposed. First, an improved generative adversarial network that incorporates a detection model and an information entropy loss function for target augmentation is proposed to suppress the generation of non-specific classes and increase the proportion of specific classes in the generated samples. Second, a multi-teacher distillation method is proposed, which employs the base and old detection models to constrain the incremental training process, thereby guiding the new detection model to detect the new specific classes and alleviate the catastrophic forgetting of old specific classes. Finally, a cost-sensitive-based gradient weighted loss is designed to guide the model focus on specific classes, and a soft-label incremental distillation loss is leveraged to enhance the generalization capability on old specific classes. The comparative experimental results with other incremental intrusion detection methods show that the average recall rate of the proposed mechanism in detecting specific classes is 51.45% and 60.32% higher than those of the compared methods after incremental training. The proposed mechanism can be adapted to existing incremental intrusion detection systems to introduce cost-sensitive, thereby improving the detection accuracy for specific attack classes and mitigating the catastrophic forgetting during the incremental training.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
    • Improved vehicle tracking framework optimized by dingo algorithm

      2026, 58(9):27-36. DOI: 10.11918/202411094

      Abstract (164) HTML (19) PDF 11.26 M (49) Comment (0) Favorites

      Abstract:In order to enhance the accuracy of vehicle target tracking in road surveillance, especially addressing the issue of reduced tracking precision and target loss due to changes in distance and occlusion, a vehicle target tracking framework based on an improved dingo optimization algorithm is proposed. Firstly, the framework integrates fast directional gradient histogram features and deep features to enhance the representation ability of the target vehicle. Secondly, by combining the Kalman filter initialization and a spiral update mechanism, the dingo optimization algorithm is improved to adapt to the problem of target loss during the search process. Finally, the improved dingo optimization algorithm is used as the search strategy to quickly locate the target, and the earth mover′s distance is selected as the similarity measurement function to maximize visual similarity. The experimental results on the UAV20L target tracking dataset show that, compared with existing tracking algorithms, the precision rate and success rate of this algorithm are both higher. Specifically, in the car1 video, the precision rate and success rate of this algorithm reached 82.9%and 22.7%, respectively; in the car16 video, the precision rate and success rate were 51.4% and 16.1%, respectively. This indicates that the algorithm can effectively deal with target drift caused by vehicle occlusion and rapid movement.The proposed tracking framework has a significant effect in improving the accuracy of vehicle target tracking and has important value for the practical application of road surveillance systems.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
    • A single-image desnowing model with joint foreground-background learning

      2026, 58(9):37-46. DOI: 10.11918/202410026

      Abstract (164) HTML (29) PDF 12.99 M (37) Comment (0) Favorites

      Abstract:To address the issue that current single-image desnowing models generally only focus on the modeling and removal of foreground snowflakes while neglecting image background information, resulting in low overall restoration accuracy, a single-image snow removal model based on joint foreground-background learning is proposed. Firstly, a dual-scale phantom feature module based on depth wise convolution is designed to enhance the representation ability of snowflakes and background details while effectively mitigating the risk of overfitting; Secondly, a joint foreground-background learning mechanism is proposed to introduce background semantic information into the snowflake modeling process, guiding the model to comprehensively understand the image structure and scene context; Finally, a bilateral long-short-term feature fusion unit is constructed. To promote full fusion and interaction between foreground and background features,which further improves the quality and stability of image restoration. Effectiveness was evaluated on two public benchmark datasets Snow100K and CSD. The results show that the proposed model outperforms existing mainstream models in terms of RPSN, SSSIM, demonstrating stronger robustness, generalization ability, and ability to process complex snow scene images. This model effectively compensates for the background modeling deficiencies in existing models and has high practical value for real-world scenarios.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
    • Diverse text-to-image synthesis based on word-level attention

      2026, 58(9):47-59. DOI: 10.11918/202412036

      Abstract (166) HTML (12) PDF 21.25 M (41) Comment (0) Favorites

      Abstract:To address the limitations of single-stage text-to-image generation technology, such as limited diversity of generated results and difficulty in precise control, this paper proposes a diverse text-to-image generative adversarial network framework based on word-level attention (WordGAN). Building on conditional generative adversarial networks, the framework incorporates a dual attention mechanism, a dual residual connection design, and a linear layer strategy. The dual attention mechanism consists of conditionally adaptive instance-layer normalization, a channel-attention module, and a pixel-attention module. It captures the weight distribution of network feature channels and spatial pixels by quantitatively analyzing the semantic contribution of each word in the text description. The dual residual connection design maintains the integrity of low-level visual features and supports deeper network expansion. The linear layer strategy uses linear layers to alter original image patterns, encouraging the framework to explore more secondary patterns. A series of experiments was conducted on the CUB bird, MSCOCO, and Oxford-102 benchmark datasets. The results show that WordGAN effectively achieves diversity in generated images without compromising quality and semantic consistency, and it outperforms other baseline text-to-image generation frameworks. The proposed diverse framework based on word-level attention provides a new approach for research on single-stage text-to-image generation tasks and demonstrates practical application value.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
      • 11+1
    • An incremental intrusion detection method based on contrastive learning and relational distillation

      2026, 58(9):60-69. DOI: 10.11918/202407062

      Abstract (148) HTML (30) PDF 5.84 M (40) Comment (0) Favorites

      Abstract:To solve the problems of catastrophic forgetting, classification bias towards new classes, and weak generalization ability of old classes in intrusion detection systems when learning new attack samples, an incremental intrusion detection method based on contrastive learning and relational distillation(CLRD-IIDM) is proposed. Firstly, a contrastive learning enhancement method is introduced, which constructs a dual branch network structure with the feature extraction layer, projection branch, and classification branch. The old class compensation contrastive loss is employed to cooperate with it to enhance the memory of old class samples; Secondly, the distillation method for the relationship between old and new classes is proposed, which extracts the similarity relationship between old and new classes in the old model and constructs class relationship labels. These labels are used to apply regularization constraints to the new model to alleviate the problem of classification bias towards new classes; Finally, the balanced bias example memory is designed to store samples that can effectively and comprehensively characterize the features of the old class, which maintains the models generalization ability to the old class. The experimental results on the UNSW-NB15 dataset show that the final accuracy of multiple neural network models using CLRD-IIDM is 2.11% to 9.78% better than the baseline model. Comparative experiments and ablation experiments have demonstrated the effectiveness and generalization of the proposed method. CLRD-IIDM can adapt well to complex and ever-changing real-world network environments, and has certain practical value.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
    • Small target detection algorithm for unmanned aerial vehicles based on baseline network lightweighting

      2026, 58(9):70-82. DOI: 10.11918/202412071

      Abstract (181) HTML (33) PDF 15.76 M (48) Comment (0) Favorites

      Abstract:To alleviate the conflicts of low pixel resolution, complex backgrounds, dense targets, and limited compute in unmanned aerial vehicle (UAV) edge detection for small targets, this paper proposed a lightweight baseline algorithm based on the improved YOLOv8s, namely you only look once v8 small lightweight (YOLOv8s-LW). First, the algorithm replaced the backbone network of YOLOv8s with a high-performance graphics processing unit network optimized with ghost and depthwise convolutions, reducing redundancy through parameter sharing and structural pruning while enhancing end-to-end trainability. Second, it replaced the module of cross-stage partial bottleneck with two convolutions with the ghost-convolution-augmented feature fusion module and the fast block-based efficient multi-scale attention fusion module, which introduced spatial mixing and multi-scale attention weighting to strengthen multi-scale feature aggregation and generalization. Subsequently, the algorithm employed a focaler-shape-wise intersection over union loss function instead of the original loss function to improve the robustness of small target regression from the perspectives of shape constraint and matching accuracy. Finally, it integrated a dynamic detection head to enhance small target responses and localization consistency under complex backgrounds. Experiments on the VisDrone-2019 dataset show that YOLOv8s-LW achieves 38.5% mAP@0.5 with 6.4×106 parameters, 19.2×109 computations, and an inference speed of 64.0 frames per second. Compared with the baseline YOLOv8s, parameters and computation are reduced by about 43% and 33%, respectively, while accuracy is improved by 0.2%. The proposed algorithm achieves a favorable trade-off among computational cost, real-time performance, and accuracy for UAV edge detection for small targets and provides engineering support and a deployable lightweight baseline and transferable architectural design for robust detection and multi-task extensions in complex background scenarios.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
      • 11+1
    • Analysis of interaction mechanism between electric field and ice coating of DC composite insulator

      2026, 58(9):83-92. DOI: 10.11918/202411087

      Abstract (123) HTML (22) PDF 15.02 M (311) Comment (0) Favorites

      Abstract:To investigate the icing mechanism of composite insulators and ensure the safety of power grids and transmission lines, a coupled-field numerical model for composite insulator icing was proposed. Using the FXBW-35/70 composite insulator as the test specimen, the motion control equation of water droplets under the interaction of airflow and electric fields was established. The electric force distribution acting on droplets at different locations of the insulator surface was analyzed. The Influence of wind speed and droplet diameter on the local collision coefficient (α1) and the droplet deflection(Δs) under a DC electric field was examined. Furthermore, the overall collision coefficient (β1) of the insulator surface under different icicle distributions was calculated and analyzed. Finally, dry and wet icing growth experiments were conducted to verify the numerical results. The results show that the electric field strength decreases successively from insulator shed edge to the insulator surface and rod. For both charged and uncharged droplets, α1 reaches its maximum at the stagnation point and decreases gradually toward both sides. Increasing wind speed or droplet diameter enhances α1 for both droplet types. Charged droplets experience stronger electric field attraction, leading to a higher α1 than uncharged droplets, and the deflection due to the electric field gradually decreases from the high-voltage to the grounded end, with a maximum deflection of up to 30%. Reducing wind speed or droplet diameter, or increasing the length or number of icicles, significantly decreases β1. Under dry growth conditions, the icing morphology is more regular without an electric field, whereas under a DC electric field, the icing rate increases and the morphology becomes irregular. Under wet growth conditions, water droplets coalesce into a film, promoting icicles growth. Without an electric field, the icing rate decreases markedly, while under a DC field, icicles become curved and distorted. This study provides a technical reference for understanding the icing mechanism of composite insulators under electric fields and for developing accurate numerical simulation methods.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
      • 11+1
      • 12+1
      • 13+1
    • Generative face detection algorithm based on multi-scale fusion and receptive field enhancement

      2026, 58(9):93-103. DOI: 10.11918/202409021

      Abstract (124) HTML (26) PDF 15.62 M (45) Comment (0) Favorites

      Abstract:In order to reduce the social media fraud and false news dissemination caused by misuse of generative face technology, a generative face detection algorithm is proposed in this paper based on YOLOv8 with multi-scale fusion and receptive field enhancement(MSERF-YOLOv8). First, a depthwise convolution module is introduced to replace the convolution module in the original network, utilizing its independent convolution kernel to process the channels and maintain feature independence; second, a new C2f module is constructed by combining the expansive residual module at the higher level of the backbone network to expand the sensory field at the higher level of the backbone network; third, an efficient multi-scale attention mechanism is introduced to aggregate the output feature maps of multiple parallel sub-networks through a cross-space learning method to enrich the feature aggregation; finally, a new hybrid architecture is constructed based on the ghost module, dynamic convolution and C2f module, the grouped convolutional parameter sharing is utilized to reconstruct the lightweight detection head. Experiments were conducted on five self-made GAN-generated face datasets, and the results show that the mean average accuracy of MSERF-YOLOv8 reaches 95.14%, which is 1.25% higher compared to YOLOv8, with the number of references being only 5.30×106 and the detection speed 90.10 frames/s. Compared with the classical target detection algorithm and some improved algorithms, the proposed algorithm has certain advantages in both feature extraction and feature fusion. The proposed algorithm improves the detection accuracy and detection speed of generative faces, while saving computational resources.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
      • 11+1
    • Intrusion detection for industrial internet stacking circular convolution and head attention

      2026, 58(9):104-114. DOI: 10.11918/202311091

      Abstract (137) HTML (16) PDF 7.34 M (36) Comment (0) Favorites

      Abstract:To address the problems of massive data, the mixture of effective and redundant information in the industrial internet, and the neglect of complex correlations among network traffic by traditional intrusion detection methods and to improve the feature extraction ability and detection accuracy of intrusion detection systems in industrial internet scenarios, this paper proposed an intrusion detection method for the industrial internet stacking circular convolution autoencoder and head attention mechanism. Firstly, a convolutional autoencoder was used to extract the spatial features of traffic data and reduce the redundancy of traffic data. Meanwhile, a two-dimensional convolutional layer and a circular padding method were introduced to expand the receptive field of the convolutional kernel on the data by increasing the dimension and expanding the spatial domain of the original data, thereby improving the feature extraction ability. Secondly, the head attention mechanism was utilized to extract the correlation information among the traffic, and attention head weights were assigned according to the importance of the correlation information extracted by different attention heads to consider the most important correlation information when determining the types of network traffic. Finally, binary classification and multi-classification comparative experiments were conducted with traditional intrusion detection methods and intrusion detection methods based on traditional time-series neural networks. Experimental results show that the intrusion detection method proposed in this paper achieves an accuracy of 99.18% in the binary classification task and 87.36% in the multi-classification task, which outperforms traditional intrusion detection methods. The proposed intrusion detection method can effectively detect network intrusions in industrial internet scenarios.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
    • Speech-to-image generation method based on a single-stage framework

      2026, 58(9):115-124. DOI: 10.11918/202403057

      Abstract (99) HTML (14) PDF 20.14 M (35) Comment (0) Favorites

      Abstract:To address the challenges in multi-stage speech-to-image generation process, such as the generators lack of integration and refinement of images, as well as the increased complexity and training costs associated with training multiple adversarial networks, we propose a single-stage speech-to-image generative adversarial network framework (Single-S2IGAN). This framework builds upon conditional generative adversarial networks by introducing a visual speech fusion module (VSFM), which consists of batch normalization (BN), a pixel-attention module (MPA), a speech modulation module (MSM), and a weighted fusion module (MWF). BN is used to stabilize the training of the network; MPA models the semantic relationships between spatial pixel regions and assigns higher weights to key positions. MSM modulates visual feature maps using fine-grained linguistic cues from the speech vector; and MWF adaptively captures the semantic importance of MPA and MSM at the channel level. A series of experiments were conducted on four benchmark datasets: CUB bird, Oxford-102, Flickr8k, and Places-subset. The results indicate that Single-S2IGAN can generate high-resolution images that are semantically consistent with the input speech descriptions. Compared to both single-stage and multi-stage frameworks, Single-S2IGAN demonstrates superior performance, with levels of effectiveness approaching those of text-to-image generation methods. The proposed Single-S2IGAN framework offers a new perspective for research into speech-to-image generation tasks.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
    • Cascade iterative refinement of Dunhuang mural image inpainting

      2026, 58(9):125-134. DOI: 10.11918/202410044

      Abstract (137) HTML (28) PDF 17.86 M (49) Comment (0) Favorites

      Abstract:To address the issues of unreasonable structural inpainting and insufficiently refined image texture features in the inpainting of damaged Dunhuang mural images, this paper proposes a cascaded iterative fine-grained mural image inpainting model. This model utilizes a multi-layer residual-net (MR-Net) combined with spatial and channel attention mechanisms,and employs an iterative layernorm residual module (LRM) together with standard convolution for initial inpainting, progressively extracting both global and local feature information of the mural images. Subsequently, the dsconv layernorm teansposed-attention (DSLT-A) and the gated feature aggregation propagation module (FAPM) are combined to reconstruct the global and local texture features. A partial convolution multi-scale gating module is used to enhance missing original image feature information during the inpainting process, refining the texture information of the mural images. Experiments were conducted using artificially irregular mask inapinting, simulated mold spot mask restoration, and real damaged mural inpainting. The experiment results show that, under three objective evaluation metrics, compared with the best-performing data among four comparative models, the model proposed in this paper improves the peak signal-to-noise ratio of the inpainted images by 1.178 dB, structural similarity index by 0.010 2, and reduces mean squared loss by 0.000 9, all outperforming the other comparative models. The proposed model is capable of inpainting higher-quality mural images, providing technical support and reference for improving the inpainting level of Dunhuang mural images and promoting future research in digital inpainting of Dunhuang murals.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
    • Infrared and visible image fusion with cross-scale and cross-attention

      2026, 58(9):135-143. DOI: 10.11918/202410008

      Abstract (131) HTML (21) PDF 13.90 M (51) Comment (0) Favorites

      Abstract:To address issues such as information loss and limitations in feature fusion during infrared and visible image fusion, a cross-scale cross-attention Transformer network is proposed. First, a cross-scale visual attention module is designed to enhance feature extraction capabilities under low-light conditions, focusing on information in low-contrast regions. Second, a cross-domain constrain module is constructed to facilitate interaction and fusion of global features through nested connections. Finally, a Transformer-based fusion strategy is proposed to further enhance image fusion performance and improve detail retention. Comparative experiments with nine other advanced fusion networks showed that in terms of objective evaluation, the cross-scale cross-attention Transformer network improved 8 evaluation indicators by 10.5% to 27.8% on the TNO dataset and by 8.6% to 30.3% on the Road scene dataset; In terms of subjective evaluation, the cross-scale cross-attention Transformer network exhibits better visual performance, significantly improving the interpretability and practicality of fused images, and effectively preserving thermal target information in infrared images and detailed textures in visible images. The proposed network can fully extract source image information, generate fusion results with clearer texture details and better contrast, providing useful references for improving image fusion performance and supporting downstream tasks.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
    • MetaDAGAN-driven approach for multifunction radar and operating mode recognition

      2026, 58(9):144-153. DOI: 10.11918/202412034

      Abstract (125) HTML (21) PDF 5.75 M (49) Comment (0) Favorites

      Abstract:To solve the problems of scarce intercepted signal samples, poor quality of augmented data, and limited recognition performance of multifunction radar (MFR) systems in complex electromagnetic environments, this paper proposed a multifunction radar and operating mode recognition method based on a meta-data augmentation generative adversarial network (MetaDAGAN) model. The model performs high-quality augmentation of small-sample signals through a generative adversarial network, thereby improving the diversity and representativeness of the training samples; introduces a meta-learning discriminator with shared structure to enable feature transfer among signals from different radar emitters and adaptive modeling of decision boundaries; and constructs a multi-task joint learning framework to integrate specific radar emitter recognition and operating mode recognition, enabling the simultaneous recognition of specific MFRs and their operating modes, so as to improve the recognition accuracies for specific MFRs and their operating modes. Comparative and ablation experiments results show that, based on MetaDAGAN, the accuracies of specific MFR recognition and operating mode recognition respectively reached 98.50% and 97.24%, which were overall superior to those of the comparative models. After the Transformer module and the sample augmentation module were respectively removed from the complete-configuration model, the operating mode recognition accuracy decreased to 72.64% and 52.06%, respectively, revealing that both of them play key roles in feature modeling and generalization capability improvement. The proposed method still maintains feature representation capability and multi-task generalization performance under small-sample conditions, demonstrating good robustness, engineering adaptability, and broad application prospects.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
    • Fusion of CBAM with bidirectional recurrent gated residual network for image Copy-Move forgery detection

      2026, 58(9):154-162. DOI: 10.11918/202412014

      Abstract (132) HTML (16) PDF 7.78 M (37) Comment (0) Favorites

      Abstract:To address the issue of insensitivity to small-scale tampering detection and low detail sensitivity in convolutional neural networks for image Copy-Move forgery detection, a bidirectional recurrent gated residual image Copy-Move forgery detection network incorporating the convolutional block attention module (CBAM) is proposed. An encoder-decoder structure is constructed using two types of residual blocks to enhance the network′s sensitivity to identify small-scale tampered regions. Building on this, the CBAM attention mechanism is introduced to strengthen feature representation capabilities and adaptively focus on critical tampered regions. Bidirectional gated recurrent units (Bi-GRU) are incorporated to interact forward and backward information, couple macro and micro features, and model long-range dependencies. Finally, a skip connection mechanism is employed to facilitate multi-scale feature fusion and strengthen the network′s recognition capability. Experiments are conducted on the public datasets CASIA2.0 and CoMoFoD. On the CASIA2.0 dataset, the proposed network achieves an accuracy of 95.78% and an F1-score of 82.08%, validating its effectiveness in small-scale tampered region detection. On the CoMoFoD dataset, it achieves an accuracy of 92.36% and an F1-score of 55.72%, further confirming its generalization capability for complex tampering scenarios. The proposed network can accurately detect Copy-Move forged images, providing technical support for subsequent research in the field of tampering detection.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
    • Estimation method of heterogeneous treatment effect with multiple treatment

      2026, 58(9):163-172. DOI: 10.11918/202411088

      Abstract (203) HTML (27) PDF 3.33 M (36) Comment (0) Favorites

      Abstract:To address the limitations of existing causal heterogeneity effect estimation methods, which primarily focus on a binary treatment and fail to accommodate multiple treatment scenarios, this paper proposes an innovative two-stage multiple treatment-causal forest (MT-CF) method. This method achieves precise estimation of causal heterogeneity effects in multiple treatment settings through a two-stage analytical process. In the first stage, traditional causal forest methods are used to estimate pairwise treatment effect differences. In the second stage, these pairwise differences are extended to the multiple treatment context, enabling the estimation of the difference between each treatment and the average effect across all treatments. Two sets of simulation experiments were designed to verify the performance of the MT-CF method in different scenarios. In addition to the simulation experiments, the MT-CF method was applied to the analysis of the breast cancer data in the database of the surveillance, epidemiology, and end results (SEER) to evaluate the individualized treatment effects of postoperative adjuvant therapies. The results indicate that the MT-CF method performed well in both simulation experiments, with multiple evaluation metrics outperforming other benchmark methods. In evaluating the individualized treatment effects of postoperative adjuvant therapies, the MT-CF method can also estimate the relative effects of various treatment options. This provides a scientific basis for making individualized treatment decisions. Furthermore, the results validate the MT-CF methods practicality and reliability when applied to real-world medical data. The proposed method not only provides an effective tool for estimating multiple treatment causal effect, but also provides a new perspective and idea for solving complex causal inference problems.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
    • Efficient time series forecasting model for feature selection

      2026, 58(9):173-182. DOI: 10.11918/202411025

      Abstract (112) HTML (22) PDF 9.31 M (37) Comment (0) Favorites

      Abstract:In order to solve the problems of high computational complexity, slow inference speed and easy loss of key information prevalent in time series prediction methods, an efficient time series prediction model with feature selection (Targetformer) is proposed. The model adopts the Transformer encoder-only structure and introduces an inverted embedding to convert the input time series into independent Token; A feature-selective attention mechanism is proposed to extract the key features from the mapped time-series Tokens. Finally, a projection layer is used to map the extracted features into the final prediction results. Experimental results on six datasets show that, compared with the best baseline model, Targetformer reduces mean squared error and mean absolute error by 5.7% and 2.0%, respectively, and achieves the best prediction performance in terms of both metrics, verifying its effectiveness and practical applicability. The model provides an efficient and practical solution in the field of time series prediction.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
    • Dynamic liquid level prediction using virtual samples generated by FD-constrained transfer

      2026, 58(9):183-195. DOI: 10.11918/202410036

      Abstract (105) HTML (16) PDF 15.40 M (37) Comment (0) Favorites

      Abstract:To solve the problem that the dynamic fluid level prediction model is ineffective due to the complex and changeable oil production process and the difficulty of obtaining complete oil well production data information, a variational autoencoder-generative adversarial network (VAE-GAN) model based on fuzzy decision constraint transfer is proposed to generate virtual samples for the dynamic liquid level prediction modeling of oil wells. Firstly, the oil well data in the source domain with relatively complete information is used to generate virtual samples through a VAE-GAN based on transfer learning. Meanwhile, the maximum mean discrepancy is used to transfer the feature distribution of virtual samples toward the target domain to be measured. Secondly, to ensure that the virtual samples possess the rich information characteristics of the source domain oil wells while conforming to the production patterns of the target domain, we design a fuzzy discriminant constraint module to distinguish the differences between the virtual samples and the data in the target domain and to further constrain the generation of the virtual samples. Finally, the obtained virtual samples are mixed with a small number of real samples in the target domain to construct the dynamic liquid level prediction model. The simulation results demonstrate that the virtual samples generated by the proposed method can effectively augment the production data of the target oil well. Compared to the prediction model constructed using actual data from the target oil well, the model developed with this method achieves a 29.589% reduction in mean absolute error and a 26.386% reduction in root mean square error, while ensuring that the relative error of key oil field production parameters remains below 5%. The proposed methodology effectively tackles the issue of incomplete information in oil well production data, thus mitigating the output error associated with the dynamic fluid level prediction model.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
    • A multi-class classification model for network intrusion detection

      2026, 58(9):196-208. DOI: 10.11918/202410075

      Abstract (126) HTML (23) PDF 10.25 M (43) Comment (0) Favorites

      Abstract:To address the limitations of existing intrusion detection models, namely, low accuracy, high false positive rates in multi-class detection, and the difficulty of efficient deployment on resource-constrained edge devices such as those used in the Internet of Things (IoT), this paper proposes a deep learning model named the convolutional long short-term self-attention network (CLSA-Net). The proposed CLSA-Net first employs a cascaded structure comprising an adaptive convolutional module, a long short-term memory network, and an improved self-attention mechanism to enhance its capability in capturing diverse temporal attack features. Subsequently, the trained model is converted into the TFLite (TensorFlow Lite) format to enable efficient deployment on low-cost ESP32-S3 embedded platforms. Comprehensive comparative experiments were conducted on multiple public datasets, along with deployment tests on the ESP32-S3 platform. The experimental results demonstrate that CLSA-Net outperforms existing mainstream models on various public datasets, including UNSW-NB15 and CICIoT2023, achieving an overall detection rate and accuracy of 98.60% and 97.48%, respectively. Furthermore, in edge deployment tests, the optimized model maintained an accuracy exceeding 80% on the ESP32-S3 platform while achieving real-time detection under low power consumption. This study validates the feasibility and practical value of deploying high-performance intrusion detection systems in resource-constrained IoT edge computing environments, striking an effective balance between high-accuracy detection and low-power operation, thus providing a practical solution for edge security defense.

      • 0+1
      • 1+1
      • 2+1
      • 3+1
      • 4+1
      • 5+1
      • 6+1
      • 7+1
      • 8+1
      • 9+1
      • 10+1
      • 11+1
      • 12+1
    • Evaluation method for node importance of Tor network based on improved ELECTRE algorithm

      2026, 58(9):209-216. DOI: 10.11918/202301024

      Abstract (126) HTML (25) PDF 2.72 M (39) Comment (0) Favorites

      Abstract:To address the challenge in current Tor network topology research of rapidly identifying critical nodes for detecting weak links and hidden services, this paper proposes a node importance evaluation method for the Tor network based on an improved ELECTRE (elimination and choice translating reality) algorithm. The proposed method integrates the entropy weight method (EWM) with the ELECTRE algorithm. After normalizing the node parameter matrix and applying entropy-based weights, the method constructs the concordance and discordance matrices, derives the aggregated dominance matrix, and ranks the nodes according to their importance. The global network efficiency, natural connectivity, and largest connected subgraph size are employed to quantify the networks invulnerability. Changes in these metrics after node removal are used to compare different node importance evaluation methods. Experimental results show that after removing 15 nodes, the network resilience of our method decreased by 79.25%; compared with the five comparison methods, the decrease in network resilience is 0.31%, 0.60%, 0.68%, 3.68%, and 0.98% higher, indicating that the method proposed in this paper has higher accuracy in node importance evaluation. Moreover, it simplifies the calculation process, avoids evaluation bias caused by human factors, and improves computational efficiency. Overall, the proposed method effectively identifies critical nodes in the Tor network, highlights their role in maintaining network structural stability, and provides strong support for threat detection, target node tracking, and vulnerability analysis in Tor network security.

      • 0+1
      • 1+1
      • 2+1
      • 3+1

Current Issue


Volume , No.

Table of Contents

Archive

Volume

Issue

Most Read

Most Cited

Download Ranking