关键词: 小样本, 度量学习, AKConv, Spatial-shift, Split-Attention, 特征融合

Abstract: Power line inspection is a core task for preventing power outages,ensuring the safe and stable operation of power grids,and promoting economic development.In practical applications,challenges such as limited defect sample numbers,diverse equipment shapes,target occlusion/adhesion,and sample imbalance are often encountered.To address these issues,this paper proposes a novel two-stage object detection network,AKS²-Net,which integrates attention and metric learning improvements.This network combines multiple attention mechanisms for feature extraction and fusion,and uses metric learning for secondary scree-ning of candidate targets,enhancing the extraction and fusion of feature information for irregular,small/fuzzy,and occluded targets,and reducing the impact of few samples and sample imbalance on network performance.Specifically,1)It designs a group convolution feature extraction network AKS² block(AKConv,Spatial-shift and Split-attention Block)based on variable convolution and spatial shift,making spatial information interaction between image features and relationship learning between feature channels possible,thereby enhancing the network's ability to mine irregular and multi-scale feature information.2)It proposes a novel multi-branch attention multi-scale feature fusion(MAFF)module,which fuses multi-channel and multi-level image detail features and spatial information through atrous spatial pyramid pooling(ASPP)and mixed skip connections(MSC),thereby improving segmentation accuracy and boundary localization ability in complex scenes.3)It proposes a feature similarity calculation method based on metric learning,which re-screens candidate negative samples by calculating the similarity between negative sample features selected by the region proposal network and all support category features,and combines a threshold to correct positive samples misjudged as negative samples,reducing interference with network training and lowering the missed detection rate of small and fuzzy targets.4)It introduces the FocalLoss function in the classification loss calculation to mitigate the impact of sample imbalance on detection results.5)Based on AKS²-Net as the backbone,a two-stage fine-tuning-based object detection network suitable for small sample and imbalanced sample conditions is constructed,providing a new option for small sample object detection through the fine-tuning mechanism.A large number of experimental results show that the proposed method performs well on the power line object detection dataset,especially enhancing the network's detection ability for far small/fuzzy and occluded objects,and has significant practical value.Additionally,under similar experimental conditions using various datasets including small sample datasets,the proposed model demonstrates stronger competitiveness compared to existing object detection networks such as ResNet 50,ResNet 101,Inception ResNet,ResNeXt 101,and ResNeSt 101.

Key words: Few-shot, Metric learning, AKConv, Spatial-shift, Split-Attention, Feature fusion