Abstract:With the rapid development of power system intelligence and unattended substations, intelligent inspection robot play an increasingly important role in equipment operation state perception and safety monitoring. Aiming at the problems of missed detection and false detection in the meter detection process of the inspection robot in the substation environment with complex lighting conditions and changing target scales, this study proposes a pointer meter detection model of substation intelligent inspection robot (SRPMNet). By constructing the average pooling down sampling (ADown) module, the computational complexity is reduced and the detection accuracy of the small target meter is improved. A lightweight attention mechanism is designed to highlight key feature information and suppress irrelevant background interference. In order to further improve the accuracy and feature expression ability of pointer meter detection in complex environment of substation, this study constructs an auxiliary detection head strategy. At the same time, the wise complete intersection over union (Wise-CIoU） is proposed in the boundary frame regression process to improve the positioning accuracy and convergence stability of the model in the complex background of the substation. This study builds a substation intelligent inspection robot platform for data collection. The experimental results show that the mAP@0.5 index of the model on the self-built substation pointer meter dataset and the public industrial meter dataset reaches 91.7% and 90.4%, respectively, which is 3.0% and 5.3% higher than the baseline model. It effectively solves the problem of missed detection and false detection, and provides effective technical support for substation intelligent inspection tasks.

Abstract:For volatile multi-metric application performance time series in cloud-native environments, we propose an integrated forecasting and monitoring method and system based on an enhanced Informer. At the encoder side, context-aware reversible normalization (CA-RevIN) is introduced to mitigate distribution shifts, while multi-frequency band convolutional enhancement (MFBCEM) and long-term dependency aggregation (LTA) are used for multi-scale modeling. A generative decoder produces multi-step forecasts. On the system side, a closed-loop architecture is built for data acquisition, feature governance, inference and alerting. Peaks-over-threshold (POT) and adaptive residual thresholds enable hierarchical anomaly detection and early warning. Experiments on public datasetsand real production data show that the proposed approach outperforms LSTM, GRU and the vanilla Informer in multi-step prediction accuracy and stability, while reducing false positives and missed detections.

Abstract:With the rapid development of intelligent driving technology, driver attention monitoring has become a crucial research focus for ensuring road safety. Traditional distraction detection methods rely on discrete behavior classification, which fails to capture the continuous variation of driver attention and often neglects individual differences in physiological structures and gaze habits.To address these limitations, this paper proposes a gaze estimation-based distraction detection method that continuously identifies distraction states by evaluating the deviation of gaze Euler angles from a predefined safe attention region. The proposed model adopts a ResNeXt-Transformer hybrid architecture that integrates local detail extraction with global dependency modeling, and incorporates a feature enhancement module to strengthen task-relevant representations and improve gaze feature extraction. Furthermore, a multi-level personalized calibration mechanism performs bias compensation at both the feature and output layers, enabling rapid individual adaptation under few-shot conditions.Experimental results on the MPIIGaze dataset demonstrate that the proposed method significantly outperforms existing approaches in cross-user gaze estimation, reducing the average angular error to 2.94°. Visualization and validation in real-world driving scenarios further confirm that the method exhibits strong interpretability and high practical applicability.

Abstract:To address the challenges of low accuracy and high model complexity in tunnel blast hole recognition caused by poor lighting and high dust levels, this study proposes a lightweight blast hole detection algorithm that integrates infrared thermal imaging with deep learning technology. Using YOLOv11n as the baseline model, the lightweight StarNet network is first incorporated to enhance multi-scale feature extraction capabilities and optimize computational efficiency. Subsequently, the Star Block module is introduced to construct the C3k2_BlockStar module, improving the fusion of features at different scales. Finally, a lightweight detection head is designed to further reduce computational complexity. Experimental results demonstrate that the improved model reduces the number of parameters by 54% and the model size to 32% of the original, achieving effective recognition of tunnel infrared blast holes. This research provides a technical foundation for robotic automated charging in tunnel construction, offering significant practical value and application prospects.

Abstract:Dermoscopic image segmentation faces significant challenges due to complex lesion morphology, blurred boundaries and noise interference. This study proposes a deep learning model named MelanoFusionNet, which employs a dual-branch encoder composed of ResNet50 and an improved Mamba VSS module. A multi-scale attention fusion decoder (MSAFD) is introduced to integrate both local and global information. Within the decoder, the extended kernel grouped gate (EKGG) and the hierarchical scale-aware attention module (HSAM) enhance feature fusion and boundary precision, while the depthwise upsample refinement block (DURB) improves reconstruction efficiency. The model is trained with a weighted combination of Dice and boundary losses. Experiments on the ISIC2018 dataset demonstrate that MelanoFusionNet outperforms mainstream methods across multiple metrics and provides reliable support for computer-aided skin cancer diagnosis.

Abstract:This study develops a three-dimensional finite-element model of a dense-electrode electric capacitance tomography （ECT） sensor under a quasi-static electrostatic field. How electrode-plate axial height, radial angle and interlayer spacing affect the capacitance dynamic range and the uniformity of sensitivity are investigated, and feasible design intervals are final determined. A composite performance objective function is constructed from the image relative error, image correlation coefficient and spatial image error, and the sensor geometry is optimized via response surface methodology (RSM). Results show that the optimized 8×8 dense electrode array markedly improves oil-water two-phase flow imaging: The image correlation coefficient is increased by 9.4%, and the image relative error and spatial image error are reduced by 13.4% and 15.2% respectively. Interaction effects among structural parameters are further analyzed to ensure robustness. The optimized dense array delivers higher resolution and accuracy, demonstrating strong potential of this ECT sensor for multiphase-flow measurement.

Abstract:Aiming at the problems of difficult parameter tuning, poor anti-interference ability and low trajectory tracking accuracy of automatic weeding control systems in complex field operating environments, which lead to difficulties in mechanical seedling avoidance and weeding between soybean plants and a high seedling damage rate, a soybean inter-plant seedling avoidance and weeding control system optimized by (multi-strategyimproved whale optimization algorithm,MSWOA） is proposed. This system controls the angular velocity of the stepper motor to achieve precise seedling avoidance and weeding with the weeding knife. Firstly, a mathematical model for inter-plant seedling avoidance and weeding is established, and a fuzzy controller is combined with a linear active disturbance rejection controller (LADRC) to improve the anti-disturbance performance of the system. Secondly, the MSWOA is used to tune the key parameters of the fuzzy-linear active disturbance rejection controller (Fuzzy-LADRC). Comparative tests with benchmark functions prove that MSWOA has better optimization performance and stability. Finally, simulation experiments are conducted on the inter-plant weeding control system model to determine the optimal control method. The simulation results show that the improved algorithm-based control method exhibits faster convergence speed and higher convergence accuracy compared with other control methods, effectively enhancing the system′s control performance. Specifically, the Fuzzy-LADRC control system optimized by MSWOA reduces the disturbance recovery time by 86.0% and 10.9% respectively in contrast to the traditional PID and LADRC control systems. Additionally, by comparing the parameter tuning effects of MSWOA with other algorithms, the superiority of the improved algorithm in this control system is further verified. This control system possesses distinct advantages in antidisturbance capability and response speed, enabling it to adapt to complex field working environments and meet the control requirements for precise seedling avoidance and weeding operations.

Abstract:In cognitive radio systems, to address the problem of inaccurate noise power estimation in traditional energy detection algorithms, a spectrum sensing optimization approach based on dynamic sub-band partitioning is proposed. Two methods are designed in this study. The first method employs the fluctuation degree of the signal envelope to divide the spectrum into multiple sub-bands and introduces an adaptive threshold coefficient to form an adaptive spectrum sensing scheme. The second method utilizes a sliding window to dynamically partition sub-bands and incorporates a single optimal threshold detection mechanism, referred to as the optimal threshold method, to achieve enhanced sub-band division and spectrum sensing.Simulation results demonstrate that, compared with the conventional single optimal threshold detection, the proposed optimal threshold method achieves up to a 10% improvement in detection performance within the SNR range of -11.7 to 0 dB, and exhibits significantly higher detection probability than the adaptive spectrum sensing method over the range of -20 to 0 dB. Overall, the optimal threshold method shows superior spectrum sensing capability and robustness.

Abstract:To address the cumulative drift issue commonly encountered during long-term autonomous navigation of unmanned forklifts in warehouses, this paper presents a localization system based on the LeGO-LOAM framework, which incorporates ground-mounted Aruco markers. The system utilizes LiDAR to provide continuous odometry constraints, while the Aruco markers offer global pose anchors upon detection. These two sources of information are fused through back-end factor graph optimization. For practical engineering deployment, this paper details the marker arrangement strategy, sensor extrinsic calibration process and the method for setting graph optimization weights. Experiments conducted in a real indoor corridor environment demonstrate that the root mean square error (RMSE) of localization is reduced from 0.635 1 m to 0.272 8 m (a decrease of 57.0%). The system maintains stable performance even under partial marker occlusion. The results indicate that this solution offers significant engineering advantages, including low implementation cost, ease of deployment and strong maintainability, making it suitable for autonomous navigation tasks in warehouse logistics scenarios.

Abstract:Aiming at the problems such as the closed encoder protocol of the displacement actuator of the spliced mirror telescope and the high power consumption and large volume of the existing decoding schemes, a high-precision position feedback system based on SoC-FPGA technology is proposed to realize the open decoding of the EnDat2.2 protocol, in order to meet the requirements of miniaturization, low power consumption and high integration for domestic substitution. The PS-PL collaborative processing architecture is constructed using the Zynq-7000 platform. The PS end uses the dual-core Cortex-A9 core to complete system control and communication scheduling. The PL end implements Manchester codec and CRC-5 verification through hardware logic. Design a driver interface circuit centered on ADM1485 and ADuM1301 to enhance the anti-interference capability. And the IDELAYE2 module is utilized for dynamic timing compensation to solve the problem of data phase offset caused by physical link differences. Experiments show that the position feedback accuracy of the system reaches 368 nm. Large-scale error injection tests demonstrate that the CRC-5 verification achieves a detection success rate of 99.99% at a 0.2% error rate, while the PCB volume is reduced by approximately 31.6% and the power consumption is lowered by more than 11 times. This solution successfully realizes high-reliability decoding of the EnDat2.2 protocol, providing crucial technical support for the miniaturization, low power consumption and high integration of displacement control systems in China′s astronomical telescopes and high-precision numerical control equipment. It holds significant engineering application value and potential for independent promotion. Future work may further optimize algorithms to enhance stability under extreme conditions.

Abstract:To address the challenges of dense stacking, blurred boundaries and uneven illumination of raw material pellets in the ironmaking process, this paper proposes a cascaded particle size detection and segmentation framework based on an improved RT-DETR and MobileSAM. In the detection stage, a coordinate attention （CA） mechanism is embedded into the backbone network to enhance feature extraction capabilities. Furthermore, an efficient discriminative frequency domain-based feedforward network （EDFFN） is introduced into the AIFI encoder and a multi-scale feature modulation （MFM） fusion module is employed to replace the traditional feature fusion method. Subsequently, the generated detection boxes serve as prompts to guide MobileSAM in generating initial masks. To optimize segmentation quality, a refiner incorporating dynamic receptive fields and efficient multi-scale attention (EMA) is designed for multi-scale adaptive repair of mask edges, which is then combined with geometric fitting to achieve precise particle size measurement. Experimental results on the pellet dataset demonstrate that the proposed method achieves a recall rate of 95.7% and a mIoU of 90.1%, effectively improving segmentation accuracy and mitigating over-segmentation. This study provides a high-precision and robust solution for online particle size monitoring in the industrial ironmaking process.

Abstract:Aiming at the problem that the measurement system of the articulated arm measuring machine cannot complete complex functions such as trajectory planning and error compensation, a measurement system for the articulated arm coordinate measuring machine based on the Unity platform was developed. The functional requirements of the articulated arm measuring machine were analyzed and the corresponding algorithm model was established. The measurement system of the measuring machine was built on the Unity platform. For the measurement system, the virtual prototype display function, the synchronous movement function of the virtual machine and the actual prototype, the multi-joint synchronous trajectory planning function and the error compensation function were respectively designed. The measurement system was used to complete the measurement experiment of the standard ball diameter. The results show that the average error after compensation is 0.015 5 mm. The integration of the measurement system functions of the articulated arm measuring machine was completed, laying the foundation for the realization of the automatic measurement of the measuring machine.

Abstract:This study presents a lightweight multi-object detection and tracking system tailored for intelligent wildlife monitoring under complex field conditions. To handle background interference, occlusion and limited edge-computing resources, the proposed framework integrates an improved YOLOv8 with ByteTrack. A global attention mechanism (GAM) enhances feature focus and noise suppression, while a lightweight YOLO-based tracking backbone (L-YOLOTrack) reduces model complexity and supports real-time deployment. In the tracking stage, a structure-aware calibration loss (SCALoss) improves localization under occlusion and an enhanced extended Kalman filter (EKF) with adaptive noise adjustment strengthens trajectory continuity under nonlinear motion. Experiments on a wildlife dataset show that mAP@0.5-0.95 increases from 40.1% to 45.7%, multiple object tracking accuracy （MOTA) rises by 14.3%, and parameters decrease by 9.3%. The results demonstrate that the proposed system achieves higher detection accuracy and tracking robustness, offering an efficient and reliable solution for wildlife monitoring.

Abstract:In the automated harvesting of cluster tomatoes in greenhouse environments, rapid identification of cluster tomatoes and precise localization of picking points are critical. This work addresses the requirements for rapid cluster recognition and accurate picking point localization in automated cluster tomato harvesting within greenhouses by proposing a visual detection and localization method based on an improved YOLOv11-Pose. The method first constructs a dataset containing simultaneous annotations of cluster tomatoes and picking points; then introduces two modules GhostC3ECA and C3k2_CA at different positions within the original YOLOv11-Pose model to achieve network lightweighting and precise picking point localization, respectively. This ultimately enables end-to-end real-time detection of cluster tomatoes and accurate picking point localization. In the network′s backend, we utilize pose-aware non-maximum suppression to optimize detection of densely clustered fruits and key points localization. Experiments validated the approach on a glasshouse-collected cluster tomato dataset, achieving 96.2% recognition accuracy and 86.5% picking point localization accuracy. The model operates at 67.2 fps, meeting real-time processing demands. Verified by real-word greenhouse harvesting scenarios, the proposed cluster tomatoes recognition and picking point localization method based on an improved YOLOv11-Pose provides an efficient and reliable visual detection solution for automated harvesting robots.

Abstract:As a core electrical parameter of high-speed interconnection links, the measurement accuracy of differential impedance and common mode impedance is critical to ensuring signal integrity. However, the fixture-induced pull effect causes the measured values to deviate significantly from their true values, which has become a prevalent challenge in the high-precision testing of high-speed connectors for Ethernet systems. Most existing studies rely on single-factor analysis methods, which fail to reveal the synergistic mechanism of tolerance and electromagnetic coupling. This paper proposes a dual-factor pull model that integrates tolerance variation and coupling strength. By introducing a tolerance variation coefficient and a coupling coefficient, a quantitative method for the pull factor is established, enabling accurate quantification of the pull effect. Simulation and experimental results indicate that the proposed model can effectively characterize the synergistic interaction mechanism between tolerance variation and electromagnetic coupling. Under the strong coupling condition, the deviation rate of impedance testing for the fixture with 3% tolerance can be controlled within 16.8% after being accurately quantified by the model, and the error between the model prediction value and the measured value is within 3%. This study provides a theoretical model and experimental basis for high-precision impedance testing, and is of great value for improving the reliability of performance evaluation for high-speed connecting equipment.

Abstract:To address low accuracy in transformer fault diagnosis under small-sample conditions, we propose a method based on an improved prototypical network. First, a residual multi-layer perceptron extracts features from transformer oil chromatography data to obtain more discriminative representations. Second, we introduce a dual-layer attention-driven prototype construction mechanism: Task-conditioned attention dynamically adjusts support-set representations, and sample attention weights support samples to emphasize discriminative features. This yields more distinguishable class prototypes and addresses prototypical networks′ limitations in cross-task adaptability and intra-class discriminability. Finally, diagnosis is performed by measuring the Euclidean distance between samples and class prototypes. Experiments on the IEEE Dataport and IEC TC 10 Dataport datasets show significant improvements: On the IEEE Dataport dataset our method outperforms the prototypical network, matching network and relation network by 15.14%, 21.49% and 18.68%, respectively; on the IEC TC 10 Dataport the gains are 6.43%, 6.90% and 8.65%, respectively. The study provide new insights and references for transformer fault diagnosis under small-sample conditions.

Abstract:Against the complex detection challenges posed by the multi-scale, multi-category, and dense distribution of objects in the daily item recognition task for intelligent household service robots, an improved YOLOv10-DLC object detection algorithm is proposed: By designing a diverse branch module for feature fusion (C2f-DBB), the algorithm network′s ability to detect objects of different scales is enhanced; By adopting a combined module of fast spatial pyramid pooling and large kernel separable convolution attention (SPPF-LSKA), the algorithm network′s ability to understand the spatial structure of features is improved, and the key features of objects are highlighted; By using the content-aware feature reassembly (CARAFE) operator, the accuracy of detail reconstruction during image upsampling in the network is improved. Experimental analysis shows that the improved YOLOv10-DLC algorithm has improved the accuracy of object detection, with its mAP value reaching 94.3%, an increase of 3.6% compared with the original algorithm network model.

Abstract:To address the challenges associated with complex background textures on the inner walls of water pipelines and the difficulty in distinguishing minute cracks from irregular fouling, this paper proposes a high-precision detection algorithm based on an improved YOLOv11n. First, the bidirectional feature pyramid network (BiFPN) is utilized to reconstruct the neck network. By employing a fast normalized fusion mechanism, this approach achieves dynamic aggregation of multi-scale features, thereby enhancing the feature representation capability for minute and variable-scale targets. Second, a global context network (GCNet) module is embedded at the end of the feature fusion stage to establish pixel-level long-range dependencies. This strengthens global background understanding and effectively suppresses noise interference such as limescale. Finally, a Focaler-MPDIoU loss function is constructed by integrating geometric optimization with a dynamic focusing mechanism to improve localization accuracy and balance the training of hard and easy samples. Experimental results demonstrate that the improved algorithm achieves an mAP50 of 91.40% on a self-constructed dataset, representing an improvement of approximately 9.5% over the baseline. The proposed method significantly enhances robustness while satisfying real-time requirements, offering substantial value for engineering applications.

Abstract:To address the issues in printed circuit board(PCB)surface defect detection networks, such as low sensitivity to small defects and low detection efficiency, a PCB defect detection algorithm CID-YOLO based on the improved YOLOv11 is proposed. First, the C3k2 module is improved by integrating Inception Depthwise Convolution, which, through a multi-branch feature extraction mechanism, reduces the loss of high-level feature information of PCB surface defects caused by increased network depth and decreases the number of model parameters. Second, a dynamic activation mechanism, Dynamic Tanh, is introduced to improve the C2PSA module, enhancing the model′s nonlinear representation capability, allowing it to better adapt to defect features of different scales. Next, the CGBD downsampling module is embedded to improve the recognition capability for special PCB surface defects through an efficient context information extraction mechanism. Finally, the original classification loss function is replaced, and a dynamic weighting strategy is applied to suppress gradient bias caused by data imbalance, effectively improving model generalization performance. Experimental results show that the proposed CID-YOLO model achieves mAP50 and mAP50.95 scores of 96.0% and 52.7%, respectively, on the Peking University public PCB defect dataset, improving 2.4% and 1.6% compared to the baseline model, demonstrating the effectiveness of the proposed algorithm.

Abstract:In response to the problem that background interference leads to insufficient extraction of key features of underwater targets and the attention area cannot cover the entire target, an underwater target detection algorithm with mixed attention and dynamic feature guidance network was proposed. This method first designs a full-dimension dynamic extraction module （FDEM） to fully understand the overall features of the target. Secondly, a weighted feature concatenation module (WFCM) is designed to retain key features from the shallow to the deep layers. The CARAFE operator is used for the feature up-sampling operation, so that the key features of local areas can receive more attention. Finally, a hybrid attention mechanism (HAM) is constructed to effectively combine the channel information and pixel information of the target and retain the key features. Experimental verification shows that this method can fully extract key features, pay more attention to the overall features of the target, thereby reducing underwater background interference and further enhancing the underwater target detection performance.

Abstract:To address the problems of difficult feature extraction and low detection accuracy caused by small target scales and severe occlusion from the UAV perspective, this paper proposes an improved RT-DETR-based UAV small target detection algorithm named SwiftHawk-DETR. Firstly, the fusion path of the feature pyramid network (FPN) is reconstructed: A tiny target detection layer P2 is added to enhance fine-grained feature capture, and the redundant large target detection layer P5 is removed to simplify computations. Secondly, a deformable fast multi-scale attention network (Dfaster_net) is designed, which strengthens the capture of small target boundary details through a dynamic sampling mechanism while adopting an efficient feature extraction strategy to reduce computational overhead. Thirdly, wavelet feature upgrade (WFU) is introduced into the neck network to alleviate the distortion of edge high-frequency features caused by upsampling and Concat operations in the original model and the slimneck architecture is combined to optimize cross-scale feature fusion. Finally, a loss function based on the weighted fusion of Focaler EIoU and NWD is constructed to improve the small target modeling ability by balancing localization errors and feature distribution differences.Experimental results on the VisDrone2019 and HIT-UAV datasets show that compared with the baseline RT-DETR, the proposed algorithm increases mAP50 and mAP50:95 by 3.8% and 4.2% respectively, and reduces the number of parameters by 61%.

Abstract:Rice is one of the main grain crops in China and the pest stress on rice restricts the healthy and sustainable development of the rice industry. Therefore, it is of great significance to quickly and accurately identify the types of rice pest infestations. An improved YOLO version 8 nano (YOLOv8n) rice pest identification and detection method was proposed to address the issues of low accuracy, easy missed and false detections in rice pest detection. Firstly, the BiFPN pyramid structure is integrated for the integration of effective feature layers. And introduce an attention mechanism module based on squeeze and excitation (SE) between the backbone network and the neck network to improve the network′s feature fusion ability; introducing the Wise-IoUv2 loss function instead of the IoU loss function to reduce the regression accuracy issue of the original IoU loss function; further enhance the detection performance of the network. Extensive experiments were conducted on the rice pest dataset, and the results showed that the improved model proposed in this paper outperforms current mainstream algorithms in terms of mAP, Precision and Recall. Specifically, this method achieved 89.2% mAP, 92.5% Precision and 83.3% Recall. In contrast, the mAP of YOLOv5 is 84.1%, YOLOv8 is 85.2%, YOLOv9 is 86.4%, YOLOv10 is 85.1% and Faster R-CNN is only 59.75%. In terms of precision, our model outperforms YOLOv5 (88.4%), YOLOv8 (89.6%) and has a higher recall than all other models,demonstrating stronger object detection ability and stability. mAP has increased by 4% compared to the original YOLOv8n. Improving the YOLOv8n algorithm can meet the requirements of rice leaf pest identification and detection accuracy, while improving the detection ability of small and dense targets, thereby reducing missed and false detections. The improved YOLOv8n algorithm has certain advantages in accuracy compared to current mainstream algorithms and provides a better method for rice pest detection, which is of great significance for the prevention and control of rice pests.

Abstract:Aiming at the problems in UAV aerial images, such as complex illumination interference, low resolution of small targets with high susceptibility to missing detection and false detection and blurred target structures, a target detection method for UAV aerial images based on improved RT-DETR is proposed, denoted as IF-DETR. Firstly, a cross-stage illumination gating unit is designed in the backbone network, which enhances the model′s ability to extract targets under complex illumination environments by modeling illumination and reflection components at the feature level. Secondly, a multi-scale feature fusion structure for small targets is constructed in the neck network, effectively improving the feature extraction and recognition capabilities for small targets. In addition, a Gaussian feature enhancement module is proposed, which utilizes Gaussian prior and spatial attention mechanism to enhance target-related structural information while suppressing background noise. Finally, a Focaler-Powerful-IoU loss function is constructed to achieve more stable and accurate target localization. Experimental results show that compared with the RT-DETR model, the improved IF-DETR algorithm on the Visdrone2019 dataset achieves improvements of 3.0%, 2.3%, 3.9% and 2.0% in mAP50, mAP50:95, recall and precision respectively. It can effectively alleviate the problems of missing detection and false detection in UAV aerial image detection, improve detection performance, and has broad application prospects.

Abstract:To address the challenges in industrial surface defect detection, such as variable morphology, low contrast and the frequent missing of small targets. A novel framework for semantic segmentation is presented, which leverages an edge-guided encoder-decoder architecture and local-global Mamba (GLMamba) to realize the synergy of "edge enhancement, global-local encoding and dual-path decoding".Firstly, a local-global Mamba encoder is built. Local details and global context are collaboratively modeled through the deployment of local and global Mamba blocks at shallow and deep layers, respectively.Secondly, an edge feature extraction module (EGFM) is designed, where edge feature maps are explicitly generated by fusing Sobel gradients with multi-scale features and are subsequently enhanced by an edge attention module (EAM). To further inject edge information into the backbone network, a multi-scale boundary gating (MSBG) mechanism is introduced to dynamically modulate encoder and decoder features. The decoder employs a dual-path attention module, where spatial details and channel context are fused in parallel to optimize multi-scale feature fusion efficiency and semantic consistency. Experimental results on the NEU-Seg, MT-Defect and FSSD-12 datasets demonstrate the effectiveness and strong generalization capability of the proposed model.

Abstract:Automatic detection of fabric defects is a critical step for quality assurance in the textile industry. However, current methods often miss defects because of difficult feature extraction. They can also misclassify defects that have little contrast with the fabric background.To address these challenges, we proposed an improved algorithm based on the YOLO version 11 nano model. First, we introduced a parameterized block with dilated convolutions into the main network backbone to enhance its ability to represent features at multiple scales. We also developed a new module to improve small object detection and model the global context. In the network′s neck, we designed a specialized attention module. This module strengthens the capture of defect shapes and textures, which reduces confusion between defects and the background. Finally, we refined the downsampling process using a re-parameterized stem structure. This change expanded the model′s receptive field and added more paths for feature extraction.Experimental results show that our improved algorithm achieves 93.2% precision, 90% recall, and 94.1% mean average precision. These results are 1.3%, 7.53% and 4.21% higher than the original model, respectively. The proposed algorithm effectively improves the accuracy of fabric defect detection and meets the practical demands of industrial production.