<p>UAV inspection of insulators is critical for power system safety, but image blur from flight jitter often obscures small-target defects. Existing models struggle with these degraded images, being either inaccurate or too computationally expensive. This necessitates a model that can efficiently and precisely detect small insulator defects in complex conditions. To solve this, we propose Sparse and Asymmetric Perception Transformer for Lightweight Insulator Defect Detection in Complex Scenes (SAP-DETR), a real-time, Transformer-based model. Its novel Focal Flow Block backbone uses sparse partial convolutions to reduce computational cost and convolutional self-attention to improve detection. In the encoder, our Multi-scale Fusion Recalibration Module weights and fuses multi-level features for better expression and contextual awareness. The Spatial Shift Aggregation Module enhances local feature perception to better capture small targets. Finally, PSConv is used for downsampling, preserving key features and capturing multi-directional information through asymmetric padding to create more expressive feature maps. Experiments demonstrate that SAP-DETR achieves high performance on the insulator defect detection dataset, with <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$AP_{50}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mn>50</mn> </msub> </mrow> </math></EquationSource> </InlineEquation> and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$AP_{50:95}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mi>A</mi> <msub> <mi>P</mi> <mrow> <mn>50</mn> <mo>:</mo> <mn>95</mn> </mrow> </msub> </mrow> </math></EquationSource> </InlineEquation> reaching 89.3% and 66.3%, respectively. Concurrently, the number of parameters and GFLOPs are reduced by 26.8% and 19.2% compared to the baseline model. In industrial applications, SAP-DETR prioritizes practicality, adaptability, and lightweight design, providing an effective and lightweight solution for insulator defect detection in complex scenes.</p>

Sparse and asymmetric perception transformer for lightweight insulator defect detection in complex scenes

期刊论文
发表日期：

2026年1月22日

查看全文

Changsheng Zhu,
Lulu Zhu,
Hongwei Bai,
Tianyu Li,
Hui Chen,
Yongxin Wang

摘要

UAV inspection of insulators is critical for power system safety, but image blur from flight jitter often obscures small-target defects. Existing models struggle with these degraded images, being either inaccurate or too computationally expensive. This necessitates a model that can efficiently and precisely detect small insulator defects in complex conditions. To solve this, we propose Sparse and Asymmetric Perception Transformer for Lightweight Insulator Defect Detection in Complex Scenes (SAP-DETR), a real-time, Transformer-based model. Its novel Focal Flow Block backbone uses sparse partial convolutions to reduce computational cost and convolutional self-attention to improve detection. In the encoder, our Multi-scale Fusion Recalibration Module weights and fuses multi-level features for better expression and contextual awareness. The Spatial Shift Aggregation Module enhances local feature perception to better capture small targets. Finally, PSConv is used for downsampling, preserving key features and capturing multi-directional information through asymmetric padding to create more expressive feature maps. Experiments demonstrate that SAP-DETR achieves high performance on the insulator defect detection dataset, with $AP_{50}$ $A P_{50}$ and $AP_{50:95}$ $A P_{50 : 95}$ reaching 89.3% and 66.3%, respectively. Concurrently, the number of parameters and GFLOPs are reduced by 26.8% and 19.2% compared to the baseline model. In industrial applications, SAP-DETR prioritizes practicality, adaptability, and lightweight design, providing an effective and lightweight solution for insulator defect detection in complex scenes.