To address the challenges of high model complexity and low accuracy in insulator component defect detection from drone-captured images,this paper presents adaptive downsampling and frequency-position fusion(ADFP),a li...To address the challenges of high model complexity and low accuracy in insulator component defect detection from drone-captured images,this paper presents adaptive downsampling and frequency-position fusion(ADFP),a lightweight algorithm based on YOLOv11(You Only Look Once version 11).The algorithm presents efficient downsampling module,new feature extraction module and innovative neck structure.By integrating the spatial channel attention module of frequency-aware cascade attention(FCA)and the ADown module,the number of parameters is reduced while accuracy is significantly improved.Additionally,the neck module is redesigned,and the position-aware key feature fusion network(PKFN)module is introduced to further improve feature fusion capabilities.Experiments were conducted on the SAID dataset using the improved model.Compared to the original model,the m AP(0.5)of ADFP shows a 5.3%improvement,while the model parameters are reduced by 12.0%.On other public insulator component defect datasets,these improvements still have better results.Multiple experiments have confirmed the effectiveness of the model and its strong generalization ability.展开更多
基金Supported by the Natural Science Foundution of Heilongjiang Province(LH2024E109)。
文摘To address the challenges of high model complexity and low accuracy in insulator component defect detection from drone-captured images,this paper presents adaptive downsampling and frequency-position fusion(ADFP),a lightweight algorithm based on YOLOv11(You Only Look Once version 11).The algorithm presents efficient downsampling module,new feature extraction module and innovative neck structure.By integrating the spatial channel attention module of frequency-aware cascade attention(FCA)and the ADown module,the number of parameters is reduced while accuracy is significantly improved.Additionally,the neck module is redesigned,and the position-aware key feature fusion network(PKFN)module is introduced to further improve feature fusion capabilities.Experiments were conducted on the SAID dataset using the improved model.Compared to the original model,the m AP(0.5)of ADFP shows a 5.3%improvement,while the model parameters are reduced by 12.0%.On other public insulator component defect datasets,these improvements still have better results.Multiple experiments have confirmed the effectiveness of the model and its strong generalization ability.
