Tianyu Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Tianyu Li has a total of 45 co-authors. Here are the top co-authors by number of shared publications:
- Changsheng Zhu
Lanzhou University of Technology
20 shared publications Hongwei Bai
Longnan Power Supply Company State Grid Gansu Power Company
9 shared publications- Jingjie Li
School of Computer and Artificial Intelligence, Lanzhou University of Technology
7 shared publications - Hui Chen
School of Computer Science and Artificial Intelligence, Lanzhou University of Technology
6 shared publications Yongxin Wang
Xi’an Jiaotong University College of Mechanical Engineering
5 shared publicationsXiaowei Feng
Gansu Tongxing Intelligent Technology Development Co Ltd
4 shared publications
Latest Publications
- 28 October 2026
- Published by IOP Publishing in Engineering Research Express
RSDMF-DETR: A Remote Sensing Direction-aware Multi-scale Fusion Detection Transformer
With the expansion of wind power scale, UAV-based remote sensing technology has become an important approach for wind turbine object detection in intelligent inspection, operation and maintenance, and wind energy resource assessment. Compared with traditional manual inspections, UAV remote sensing provides advantages of non-contact operation, wide coverage, rapid acquisition, and high spatial resolution. However, wind turbine objects exhibit significant scale variations, diverse orientations, and complex backgrounds, posing challenges to feature representation, multi-scale fusion, and robust detection. To address these challenges, this paper constructs a UAV-based remote sensing dataset and proposes a Remote Sensing Direction-aware Multi-scale Fusion Detection Transformer (RSDMF-DETR), a task-oriented, stage-wise framework that integrates existing feature extraction components with two self-designed modules, Remote Spatial Direction Attention (RSDA) and Scale-Enhanced Bidirectional Channel-Spatial Attention (SE-BCSA). Specifically, the backbone adopts the existing Context-Guided Block (CGB) to jointly model local texture information and contextual semantic cues, thereby improving feature representation under complex backgrounds. The Remote Direction-Aware Encoding Network (RDEN) integrates existing GSConv and VoVGSCSP components with the proposed RSDA module to enhance feature encoding, cross-layer aggregation, and directional structural representation. Finally, the Remote Multi-scale Adaptive Fusion (RMAF) stage combines the existing DynamicScalSeq component with the proposed SE-BCSA module to enhance cross-scale feature alignment and channel-spatial feature refinement. Experimental results on the self-built UAV remote sensing dataset demonstrate that RSDMF-DETR achieves 89.5% 〖AP〗_(IoU=50) and 67.1% 〖AP〗_(IoU=50:95), representing improvements of 2.0 and 0.8 percentage points over RT-DETR-R18, respectively. These results demonstrate that the proposed stage-wise strategy effectively integrates contextual, directional, and multi-scale feature representations for wind turbine detection under complex remote sensing conditions. RSDMF-DETR may provide target-level information for wind farm inspection, localization, and spatial information management. Its deployment and performance on actual UAV or edge-computing platforms have not yet been experimentally validated and remain topics for future work.
- 28 October 2026
- Published by IOP Publishing in Engineering Research Express
RSDMF-DETR: A Remote Sensing Direction-aware Multi-scale Fusion Detection Transformer
With the expansion of wind power scale, UAV-based remote sensing technology has become an important approach for wind turbine object detection in intelligent inspection, operation and maintenance, and wind energy resource assessment. Compared with traditional manual inspections, UAV remote sensing provides advantages of non-contact operation, wide coverage, rapid acquisition, and high spatial resolution. However, wind turbine objects exhibit significant scale variations, diverse orientations, and complex backgrounds, posing challenges to feature representation, multi-scale fusion, and robust detection. To address these challenges, this paper constructs a UAV-based remote sensing dataset and proposes a Remote Sensing Direction-aware Multi-scale Fusion Detection Transformer (RSDMF-DETR), a task-oriented, stage-wise framework that integrates existing feature extraction components with two self-designed modules, Remote Spatial Direction Attention (RSDA) and Scale-Enhanced Bidirectional Channel-Spatial Attention (SE-BCSA). Specifically, the backbone adopts the existing Context-Guided Block (CGB) to jointly model local texture information and contextual semantic cues, thereby improving feature representation under complex backgrounds. The Remote Direction-Aware Encoding Network (RDEN) integrates existing GSConv and VoVGSCSP components with the proposed RSDA module to enhance feature encoding, cross-layer aggregation, and directional structural representation. Finally, the Remote Multi-scale Adaptive Fusion (RMAF) stage combines the existing DynamicScalSeq component with the proposed SE-BCSA module to enhance cross-scale feature alignment and channel-spatial feature refinement. Experimental results on the self-built UAV remote sensing dataset demonstrate that RSDMF-DETR achieves 89.5% 〖AP〗_(IoU=50) and 67.1% 〖AP〗_(IoU=50:95), representing improvements of 2.0 and 0.8 percentage points over RT-DETR-R18, respectively. These results demonstrate that the proposed stage-wise strategy effectively integrates contextual, directional, and multi-scale feature representations for wind turbine detection under complex remote sensing conditions. RSDMF-DETR may provide target-level information for wind farm inspection, localization, and spatial information management. Its deployment and performance on actual UAV or edge-computing platforms have not yet been experimentally validated and remain topics for future work.
- 1 October 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (20) , 205215
A data-efficient framework for power line inspection using semantic-guided synthesis and lightweight HypMamba-DT
Suspended foreign objects on power lines pose significant risks to grid stability, yet their automated detection is often hindered by the scarcity of anomalous samples and the high computational demands of deployment. To enhance model robustness against environmental variability, this study first utilizes an artificial intelligence generated content-based generative pipeline to synthesize high-fidelity datasets representing rare meteorological conditions, establishing a robust foundation for training. Building upon this, we propose HypMamba-DT, a lightweight and high-performance detection architecture that innovatively integrates Mamba’s selective state-space modeling with hypergraph theory. By optimizing cross-scale feature fusion, the model achieves superior discriminative precision for small targets with a significantly reduced overhead of only 6 M parameters and 12.89 GFLOPs. Furthermore, we implement a cloud-edge-client collaborative framework for practical industrial application. As demonstrated in embedded hardware deployment evaluation, the system balances local edge sensing on embedded devices with cloud-based diagnostics. Experimental results on both public and high-definition datasets confirm that the proposed system satisfies stringent industrial requirements, achieving a real-time inference speed of 46 FPS in cloud-assisted mode and 25 FPS on edge platforms. This integrated solution provides a highly efficient and scalable approach for the automated maintenance of modern power systems. The code has been demonstrated in https://github.com/Tianyu-Li-S/HypMamba-DT .
- 1 October 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (20) , 205215
A data-efficient framework for power line inspection using semantic-guided synthesis and lightweight HypMamba-DT
Suspended foreign objects on power lines pose significant risks to grid stability, yet their automated detection is often hindered by the scarcity of anomalous samples and the high computational demands of deployment. To enhance model robustness against environmental variability, this study first utilizes an artificial intelligence generated content-based generative pipeline to synthesize high-fidelity datasets representing rare meteorological conditions, establishing a robust foundation for training. Building upon this, we propose HypMamba-DT, a lightweight and high-performance detection architecture that innovatively integrates Mamba’s selective state-space modeling with hypergraph theory. By optimizing cross-scale feature fusion, the model achieves superior discriminative precision for small targets with a significantly reduced overhead of only 6 M parameters and 12.89 GFLOPs. Furthermore, we implement a cloud-edge-client collaborative framework for practical industrial application. As demonstrated in embedded hardware deployment evaluation, the system balances local edge sensing on embedded devices with cloud-based diagnostics. Experimental results on both public and high-definition datasets confirm that the proposed system satisfies stringent industrial requirements, achieving a real-time inference speed of 46 FPS in cloud-assisted mode and 25 FPS on edge platforms. This integrated solution provides a highly efficient and scalable approach for the automated maintenance of modern power systems. The code has been demonstrated in https://github.com/Tianyu-Li-S/HypMamba-DT .
- 1 September 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (18) , 185406
Intrusion detection based on the ManFormer model
Network intrusion detection is a key component for ensuring cyberspace security and stable network operation. To address the limitations of traditional methods in high-dimensional network traffic, such as insufficient representation capability, degraded performance in classifying complex attacks, and limited generalization ability, this study designs and implements ManFormer, a hybrid Transformer–Mamba classification architecture that integrates Transformer and Mamba for high-dimensional imbalanced classification tasks. The proposed model uses a Transformer module to capture global interactions among PCA-derived latent feature tokens, and further introduces a Mamba state-space module to refine the dependencies within the context-enhanced latent feature sequence. This design improves, to some extent, the comprehensive recognition capability for complex attack categories and improves Macro- F 1 to a certain extent. During data preprocessing, Borderline-SMOTE is applied to oversample minority-class samples in the training set, with a focus on difficult samples near class boundaries, thereby alleviating the class imbalance problem. In addition, CBLoss is adopted as the training loss function to dynamically adjust the weights of different classes and further compensate for differences in sample size. Experimental results show that ManFormer achieves good overall performance in high-dimensional imbalanced multi-class intrusion detection tasks. Although it improves Macro- F 1 compared with standalone Transformer and Mamba models, the stable recognition of extremely rare attack classes remains challenging. Therefore, the proposed method should be regarded as an effective mitigation strategy for imbalanced intrusion detection, rather than a complete solution to extreme rare-class recognition.
- 1 September 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (18) , 185406
Intrusion detection based on the ManFormer model
Network intrusion detection is a key component for ensuring cyberspace security and stable network operation. To address the limitations of traditional methods in high-dimensional network traffic, such as insufficient representation capability, degraded performance in classifying complex attacks, and limited generalization ability, this study designs and implements ManFormer, a hybrid Transformer–Mamba classification architecture that integrates Transformer and Mamba for high-dimensional imbalanced classification tasks. The proposed model uses a Transformer module to capture global interactions among PCA-derived latent feature tokens, and further introduces a Mamba state-space module to refine the dependencies within the context-enhanced latent feature sequence. This design improves, to some extent, the comprehensive recognition capability for complex attack categories and improves Macro- F 1 to a certain extent. During data preprocessing, Borderline-SMOTE is applied to oversample minority-class samples in the training set, with a focus on difficult samples near class boundaries, thereby alleviating the class imbalance problem. In addition, CBLoss is adopted as the training loss function to dynamically adjust the weights of different classes and further compensate for differences in sample size. Experimental results show that ManFormer achieves good overall performance in high-dimensional imbalanced multi-class intrusion detection tasks. Although it improves Macro- F 1 compared with standalone Transformer and Mamba models, the stable recognition of extremely rare attack classes remains challenging. Therefore, the proposed method should be regarded as an effective mitigation strategy for imbalanced intrusion detection, rather than a complete solution to extreme rare-class recognition.
- 14 August 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (33) , 335902
Research on the prediction of coupled growth of alloy solidification dendrites and gas bubbles based on the MLLAGP model
Accurately simulating the complex spatiotemporal dynamics of dendritic growth and gas bubble interactions during alloy solidification is crucial for predicting final material properties and controlling defects. Conventional numerical methods, such as the coupled cellular automata (CA)-finite difference method (FDM)-lattice Boltzmann method (LBM) framework, can resolve the mechanisms of dendritic evolution and bubble formation; however, their prohibitive computational costs restrict their application in long-term, large-scale material design. To address this challenge, this study proposes Mamba-like Linear Attention-based growth predictor (MLLAGP), a deep learning-based spatiotemporal prediction model centered on a Linear Attention mechanism. By integrating a multi-scale encoder with a specialized MLLA Block-which incorporates conditional positional encoding and depthwise convolutions-the model achieves linear computational complexity while simultaneously capturing local microstructural details and global spatiotemporal dependencies. This architecture enables the precise modeling of key physical features, such as the competitive growth of bubbles within interdendritic regions and their dynamic interactions with the solid–liquid interface. High-fidelity microstructural data generated via the CA-FDM-LBM framework were utilized to analyze the spatiotemporal coupling between dendrites and bubbles, facilitating the training of a nonlinear mapping from initial states to future evolutionary trajectories. Experimental results demonstrate that MLLAGP successfully overcomes the computational bottlenecks of standard Transformer architectures when processing high-resolution physical fields. Compared with other state-of-the-art models, MLLAGP achieves superior predictive accuracy in both spinodal decomposition and alloy solidification tasks while maintaining a significantly lower parameter count. Notably, in physical fidelity evaluations, the predicted area fractions for dendrites and bubbles stabilized at ratios of 1.01 and 0.85–1.05 relative to ground truth, respectively, indicating statistical consistency with the mass conservation trends of numerical simulations. MLLAGP not only dramatically enhances computational efficiency but also provides a robust tool for accelerating new material development and optimizing solidification processes.
- 14 August 2026
- Published by IOP Publishing in Physica Scripta
- Vol. 101 (33) , 335902
Research on the prediction of coupled growth of alloy solidification dendrites and gas bubbles based on the MLLAGP model
Accurately simulating the complex spatiotemporal dynamics of dendritic growth and gas bubble interactions during alloy solidification is crucial for predicting final material properties and controlling defects. Conventional numerical methods, such as the coupled cellular automata (CA)-finite difference method (FDM)-lattice Boltzmann method (LBM) framework, can resolve the mechanisms of dendritic evolution and bubble formation; however, their prohibitive computational costs restrict their application in long-term, large-scale material design. To address this challenge, this study proposes Mamba-like Linear Attention-based growth predictor (MLLAGP), a deep learning-based spatiotemporal prediction model centered on a Linear Attention mechanism. By integrating a multi-scale encoder with a specialized MLLA Block-which incorporates conditional positional encoding and depthwise convolutions-the model achieves linear computational complexity while simultaneously capturing local microstructural details and global spatiotemporal dependencies. This architecture enables the precise modeling of key physical features, such as the competitive growth of bubbles within interdendritic regions and their dynamic interactions with the solid–liquid interface. High-fidelity microstructural data generated via the CA-FDM-LBM framework were utilized to analyze the spatiotemporal coupling between dendrites and bubbles, facilitating the training of a nonlinear mapping from initial states to future evolutionary trajectories. Experimental results demonstrate that MLLAGP successfully overcomes the computational bottlenecks of standard Transformer architectures when processing high-resolution physical fields. Compared with other state-of-the-art models, MLLAGP achieves superior predictive accuracy in both spinodal decomposition and alloy solidification tasks while maintaining a significantly lower parameter count. Notably, in physical fidelity evaluations, the predicted area fractions for dendrites and bubbles stabilized at ratios of 1.01 and 0.85–1.05 relative to ground truth, respectively, indicating statistical consistency with the mass conservation trends of numerical simulations. MLLAGP not only dramatically enhances computational efficiency but also provides a robust tool for accelerating new material development and optimizing solidification processes.
- 8 July 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (27) , 275408
LAF-DETR: lightweight attentive fusion DETR for small and occluded object detection in aerial images
Small object detection in unmanned aerial vehicle aerial images presents significant challenges because objects are tiny, feature information is limited, and dense scenes often contain partial occlusion. Existing detection algorithms often struggle to balance accuracy and efficiency under these conditions. To address these problems, this paper proposes a lightweight attention fusion model based on DETR, termed LAF-DETR. The model adopts a task-oriented redesign of the backbone, neck, and detection head. The lightweight attention network backbone reduces computational redundancy while improving feature discriminability. The multi-scale neck combines content-aware upsampling, wavelet-based downsampling, and lightweight grouped fusion to preserve fine-grained details during feature propagation. The detection head integrates S2 high-resolution feature maps and an attentional scale sequence fusion mechanism to strengthen cross-scale feature representation for small and partially occluded objects. Experiments on VisDrone2019 show that, compared with RT-DETR-R18, LAF-DETR improves mAP 50 : 95 by 2.6% while reducing the parameter count and GFLOPs by 50.8% and 30.9%, respectively. Practical speed measurements further show that LAF-DETR reaches 90 FPS on the stated hardware. Additional horizontal bounding box evaluation on DOTAv1.0 confirms the generalization ability of the proposed method in aerial small object detection.
- 8 July 2026
- Published by IOP Publishing in Measurement Science and Technology
- Vol. 37 (27) , 275408
LAF-DETR: lightweight attentive fusion DETR for small and occluded object detection in aerial images
Small object detection in unmanned aerial vehicle aerial images presents significant challenges because objects are tiny, feature information is limited, and dense scenes often contain partial occlusion. Existing detection algorithms often struggle to balance accuracy and efficiency under these conditions. To address these problems, this paper proposes a lightweight attention fusion model based on DETR, termed LAF-DETR. The model adopts a task-oriented redesign of the backbone, neck, and detection head. The lightweight attention network backbone reduces computational redundancy while improving feature discriminability. The multi-scale neck combines content-aware upsampling, wavelet-based downsampling, and lightweight grouped fusion to preserve fine-grained details during feature propagation. The detection head integrates S2 high-resolution feature maps and an attentional scale sequence fusion mechanism to strengthen cross-scale feature representation for small and partially occluded objects. Experiments on VisDrone2019 show that, compared with RT-DETR-R18, LAF-DETR improves mAP 50 : 95 by 2.6% while reducing the parameter count and GFLOPs by 50.8% and 30.9%, respectively. Practical speed measurements further show that LAF-DETR reaches 90 FPS on the stated hardware. Additional horizontal bounding box evaluation on DOTAv1.0 confirms the generalization ability of the proposed method in aerial small object detection.
- 1 July 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13) , 135345
CFENet: a real-time detection model suitable for infrared fault detection in electrical equipment
Infrared detection plays a critical role in identifying faults in electrical equipment. However, infrared images often suffer from blurred target boundaries, complex background thermal noise, low contrast, and indistinct small objects. These issues make it challenging for existing algorithms to balance detection accuracy and real-time performance. To address these limitations, this paper proposes a lightweight real-time detection model, CFENet (context feature extraction network). The core contributions of CFENet are threefold: First, in the backbone network, the IFAE (infrared feature anti-interference enhancement) and MFDown (multi-scale fidelity-preserving downsampling) modules are introduced to establish multi-scale contextual extraction capability. The IFAE module employs lightweight attention to enhance thermal anomaly responses, while the MFDown suppresses background noise during the downsampling process, thereby collectively improving the semantic saliency of fault regions. Second, the neck network incorporates a dynamic adaptive architecture that integrates the CFConcat (context feature concatenation) and EFC (efficient feature cleaning) modules for multi-layer semantic fusion. By leveraging the RCM (rectangular self-calibration module), the model establishes cross-scale contextual correlations, thereby enhancing feature coherence for blurred boundaries and small targets. Finally, at the detection head level, the detection head employs DyHead (dynamic detection head), which uses a channel-position synergistic enhancement mechanism to accurately discriminate weak fault features from complex background noise. To evaluate the performance of CFENet, experiments were conducted on two infrared fault datasets: Infrared image of power equipment and high voltage defect detection in electrical equipment. The results demonstrate that CFENet-N achieves mAP50 improvements of 7.5% and 3.9% over YOLOv8-N, while CFENet-S surpasses YOLOv6-S by 8.3% and 3.8%, respectively. Both models maintain high inference speeds while significantly improving detection accuracy, offering an efficient solution for real-time online monitoring and edge deployment in infrared fault detection of power equipment. Code is available at https://github.com/ttLovelace/CFENet .
- 1 July 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13) , 135345
CFENet: a real-time detection model suitable for infrared fault detection in electrical equipment
Infrared detection plays a critical role in identifying faults in electrical equipment. However, infrared images often suffer from blurred target boundaries, complex background thermal noise, low contrast, and indistinct small objects. These issues make it challenging for existing algorithms to balance detection accuracy and real-time performance. To address these limitations, this paper proposes a lightweight real-time detection model, CFENet (context feature extraction network). The core contributions of CFENet are threefold: First, in the backbone network, the IFAE (infrared feature anti-interference enhancement) and MFDown (multi-scale fidelity-preserving downsampling) modules are introduced to establish multi-scale contextual extraction capability. The IFAE module employs lightweight attention to enhance thermal anomaly responses, while the MFDown suppresses background noise during the downsampling process, thereby collectively improving the semantic saliency of fault regions. Second, the neck network incorporates a dynamic adaptive architecture that integrates the CFConcat (context feature concatenation) and EFC (efficient feature cleaning) modules for multi-layer semantic fusion. By leveraging the RCM (rectangular self-calibration module), the model establishes cross-scale contextual correlations, thereby enhancing feature coherence for blurred boundaries and small targets. Finally, at the detection head level, the detection head employs DyHead (dynamic detection head), which uses a channel-position synergistic enhancement mechanism to accurately discriminate weak fault features from complex background noise. To evaluate the performance of CFENet, experiments were conducted on two infrared fault datasets: Infrared image of power equipment and high voltage defect detection in electrical equipment. The results demonstrate that CFENet-N achieves mAP50 improvements of 7.5% and 3.9% over YOLOv8-N, while CFENet-S surpasses YOLOv6-S by 8.3% and 3.8%, respectively. Both models maintain high inference speeds while significantly improving detection accuracy, offering an efficient solution for real-time online monitoring and edge deployment in infrared fault detection of power equipment. Code is available at https://github.com/ttLovelace/CFENet .
- 25 June 2026
- Published by Springer Nature in Machine Vision and Applications
- Vol. 37 (5) , 102
Multiscale feature optimization for accurate small object detection in remote sensing imagery
- 25 June 2026
- Published by Springer Nature in Machine Vision and Applications
- Vol. 37 (5) , 102
Multiscale feature optimization for accurate small object detection in remote sensing imagery
- 18 June 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13)
MDCG-YOLO:Rapid Real-Time Detection Model for Suspended Foreign Objects on Power Lines
In the field of object detection, particularly in the complex domain of detecting foreign objects on overhead power lines, the core challenge lies in addressing the insufficient generalization capabilities of existing models and their difficulty in handling the diverse demands of complex scenarios. We propose an innovative lightweight, fast, real-time detection model based on the YOLO architecture. To tackle the challenge of detecting minute foreign objects on power lines, this architecture effectively reduces computational overhead and enhances feature extraction capabilities through the DSC3K2 and C3K2CG modules within its backbone network. The ContextGuided module within C3K2CG integrates local features, surrounding contextual information, and global contextual information to better capture semantic and spatial relationships within images. Furthermore, to address the poor fusion capability of Neck for feature maps at different scales and to adapt to diverse complex line scenarios, this paper introduces the ContAttn module for feature splicing. This module enhances the fusion capability for feature maps with inconsistent information representation and scale features. To address the limitation of upsampling failing to adaptively adjust based on varying feature map contents, DySample is adopted for upsampling operations, thereby strengthening the model's generalization capability. Furthermore, ADown and SCDown replace native convolutions. Through lightweight design and adaptive sampling, more information is retained while reducing computational load and parameter count. Experimental results demonstrate
- 18 June 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13)
MDCG-YOLO:Rapid Real-Time Detection Model for Suspended Foreign Objects on Power Lines
In the field of object detection, particularly in the complex domain of detecting foreign objects on overhead power lines, the core challenge lies in addressing the insufficient generalization capabilities of existing models and their difficulty in handling the diverse demands of complex scenarios. We propose an innovative lightweight, fast, real-time detection model based on the YOLO architecture. To tackle the challenge of detecting minute foreign objects on power lines, this architecture effectively reduces computational overhead and enhances feature extraction capabilities through the DSC3K2 and C3K2CG modules within its backbone network. The ContextGuided module within C3K2CG integrates local features, surrounding contextual information, and global contextual information to better capture semantic and spatial relationships within images. Furthermore, to address the poor fusion capability of Neck for feature maps at different scales and to adapt to diverse complex line scenarios, this paper introduces the ContAttn module for feature splicing. This module enhances the fusion capability for feature maps with inconsistent information representation and scale features. To address the limitation of upsampling failing to adaptively adjust based on varying feature map contents, DySample is adopted for upsampling operations, thereby strengthening the model's generalization capability. Furthermore, ADown and SCDown replace native convolutions. Through lightweight design and adaptive sampling, more information is retained while reducing computational load and parameter count. Experimental results demonstrate
- 1 April 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (9)
DCAF-DETR: Content-Adaptive Resampling with Scale-Conditioned Selective Aggregation for fuzzy Small-Object Detection in Real-Time UAV Power-Line Inspection
Real-time detection of small, visually ambiguous targets in UAV power-line inspection presents a significant challenge due to the limited computational capacity of onboard processors. In this work, we propose DCAF-DETR, a lightweight detection framework designed to enhance the recognition of fuzzy small objects within the RT-DETR paradigm. The architecture integrates a CGRepBlock backbone and the DySample content-adaptive upsampling module to preserve fine visual details, alongside contextual modules to maintain semantic stability. To address feature aliasing, three scale-discriminative fusion components-Fusion-CADA, Fusion-DCPA, and Fusion-SDFM-are introduced. To optimize the accuracy-efficiency trade-off, DCAF-DETR prioritizes high inference speed and low resource consumption. Evaluated on public datasets, the model achieves 76.5\% $AP_{50}$ and 35.6\% $AP_{50:95}$. Although the high-precision metric is marginally lower than heavy state-of-the-art models (e.g., YOLOv8-X), our framework runs at 65 FPS with only 12.55M parameters and 40.3 GFLOPs-reducing computational cost by approximately 80\% compared to large-scale counterparts. This confirms that DCAF-DETR achieves a superior Pareto frontier suitable for resource-constrained UAV platforms. Validation on a custom dataset ($AP_{50}$ of 80.5\%) further confirms its robustness, demonstrating DCAF-DETR as a practical solution for automated inspection tasks.- 1 April 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (9)
DCAF-DETR: Content-Adaptive Resampling with Scale-Conditioned Selective Aggregation for fuzzy Small-Object Detection in Real-Time UAV Power-Line Inspection
Real-time detection of small, visually ambiguous targets in UAV power-line inspection presents a significant challenge due to the limited computational capacity of onboard processors. In this work, we propose DCAF-DETR, a lightweight detection framework designed to enhance the recognition of fuzzy small objects within the RT-DETR paradigm. The architecture integrates a CGRepBlock backbone and the DySample content-adaptive upsampling module to preserve fine visual details, alongside contextual modules to maintain semantic stability. To address feature aliasing, three scale-discriminative fusion components-Fusion-CADA, Fusion-DCPA, and Fusion-SDFM-are introduced. To optimize the accuracy-efficiency trade-off, DCAF-DETR prioritizes high inference speed and low resource consumption. Evaluated on public datasets, the model achieves 76.5\% $AP_{50}$ and 35.6\% $AP_{50:95}$. Although the high-precision metric is marginally lower than heavy state-of-the-art models (e.g., YOLOv8-X), our framework runs at 65 FPS with only 12.55M parameters and 40.3 GFLOPs-reducing computational cost by approximately 80\% compared to large-scale counterparts. This confirms that DCAF-DETR achieves a superior Pareto frontier suitable for resource-constrained UAV platforms. Validation on a custom dataset ($AP_{50}$ of 80.5\%) further confirms its robustness, demonstrating DCAF-DETR as a practical solution for automated inspection tasks.- 1 February 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (4) , 045224
MG-fusion-DETR: a lightweight transformer-based multi-granularity detection model
Against the backdrop of accelerating intelligent transformation in global energy infrastructure, establishing a secure, stable, and efficient smart monitoring system for transmission and distribution grids has become a core requirement. In response to the increasingly severe threat of foreign object intrusion in transmission lines, traditional reliance on manual inspections struggles to meet the demands of real-time proactive defense due to inherent limitations such as blind spots in mountainous areas, difficulty in identifying small targets, and delayed fault response. To overcome these challenges, this study proposes a lightweight object detection model based on the Transformer architecture—Multi-Granularity Fusion-DETR (MG-Fusion-DETR)—aiming to achieve a delicate balance between detection performance and computational efficiency. The core innovation of the model lies in two aspects: firstly, a Multi-Granularity Fusion Attention (MGFA) mechanism is proposed, which achieves precise focusing and discriminative learning for complex texture features and small-scale foreign objects by coupling pooled channel attention with an original multi-scale depthwise separable convolution structure; secondly, a bottleneck residual unit based on exponential moving average (EMA) fast convolution is designed, which significantly reduces computational overhead while preserving fine-grained feature extraction capability, laying the foundation for lightweight deployment of the model. On this basis, we construct a Multi-Granularity Context Fusion Transformer Encoder to integrate the global-local modeling advantages of MGFA, and introduce an auxiliary global context coordination module to further optimize feature fusion. Experimental results demonstrate that MG-Fusion-DETR achieves detection accuracies of 85.3% and 87.3% on public and self-built high-definition foreign object detection datasets, respectively, while reaching an inference speed of 160 FPS. This fully validates its outstanding performance and significant engineering potential for real-time foreign object detection tasks on power lines.
- 1 February 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (4) , 045224
MG-fusion-DETR: a lightweight transformer-based multi-granularity detection model
Against the backdrop of accelerating intelligent transformation in global energy infrastructure, establishing a secure, stable, and efficient smart monitoring system for transmission and distribution grids has become a core requirement. In response to the increasingly severe threat of foreign object intrusion in transmission lines, traditional reliance on manual inspections struggles to meet the demands of real-time proactive defense due to inherent limitations such as blind spots in mountainous areas, difficulty in identifying small targets, and delayed fault response. To overcome these challenges, this study proposes a lightweight object detection model based on the Transformer architecture—Multi-Granularity Fusion-DETR (MG-Fusion-DETR)—aiming to achieve a delicate balance between detection performance and computational efficiency. The core innovation of the model lies in two aspects: firstly, a Multi-Granularity Fusion Attention (MGFA) mechanism is proposed, which achieves precise focusing and discriminative learning for complex texture features and small-scale foreign objects by coupling pooled channel attention with an original multi-scale depthwise separable convolution structure; secondly, a bottleneck residual unit based on exponential moving average (EMA) fast convolution is designed, which significantly reduces computational overhead while preserving fine-grained feature extraction capability, laying the foundation for lightweight deployment of the model. On this basis, we construct a Multi-Granularity Context Fusion Transformer Encoder to integrate the global-local modeling advantages of MGFA, and introduce an auxiliary global context coordination module to further optimize feature fusion. Experimental results demonstrate that MG-Fusion-DETR achieves detection accuracies of 85.3% and 87.3% on public and self-built high-definition foreign object detection datasets, respectively, while reaching an inference speed of 160 FPS. This fully validates its outstanding performance and significant engineering potential for real-time foreign object detection tasks on power lines.