摘要
为解决电力作业环境中背景复杂和目标尺寸多样导致的穿戴检测难题,提出一种轻量化的边缘感知检测算法;采用多尺度边缘信息增强方法在浅层高效融合边缘特征以抑制背景噪声;设计ShiftRepC3模块于混合编码器中高效提取并融合局部特征,并利用DySample改进上采样增强重建能力;引入基于注意力机制的剪枝方法降低模型参数量与计算量,满足边缘设备部署需求;实验结果表明,和基线模型RT-DETR相比,所提方法在两个数据集上的mAP50分别提升了2.8%和2.1%,在mAP 50-95分别提升了2.6%和1.5%,并且保持了较低计算开销。
To solve the wearable detection of complex backgrounds and diverse target sizes in power operation environments,a lightweight edge-aware detection algorithm is proposed.A multi-scale edge information enhancement method is employed to efficiently fuse edge features at shallow layers,suppressing background noise.A ShiftRepC3 module is designed within the hybrid encoder to effectively extract and fuse local features,and a DySample method is utilized to improve upsampling for the capability of enhanced reconstruction.An attention mechanism-based pruning method is introduced to reduce the parameters and computational costs of the model,meeting the deployment requirement of edge devices.Experimental results demonstrate that,compared with the baseline real-time detection transformer(RT-DETR)model on two datasets,the proposed method improves the mAP 50 by 2.8%and 2.1%,and the mAP 50-95 by 2.6%and 1.5%,respectively,and maintains a low computational cost.
作者
柏帆
郭鹏程
佟鑫
王荣历
刘佳杰
金彬
BAI Fan;GUO Pengcheng;TONG Xin;WANG Rongli;LIU Jiajie;JIN Bin(Ninghai Yancangshan Electric Power Construction Co.,Ltd.,Ningbo 315600,China;Ninghai County Power Supply Company,State Grid Zhejiang Electric Power Co.,Ltd.,Ningbo 315600,China)
出处
《计算机测量与控制》
2025年第8期112-119,128,共9页
Computer Measurement & Control
基金
宁波永耀电力投资集团科技项目(CF058211002024001)。
