【摘要】 针对输电线路巡检（TLI）中存在的复杂背景干扰、目标尺度跨度大及器件密集分布导致小目标检测精度低的问题，提出一种适用于TLI的检测变压器（DETR）模型TLI-DETR。首先，通过在骨干网络各阶段末层用多尺度可分离卷积网络替换残差块，提取并融合各阶段输出的特征图作为多尺度特征，以增强小目标特征提取能力；其次，在多尺度特征图中引入通道空间融合交叉注意力机制，抑制图像中背景噪声对小目标检测的干扰；最后，设计层级交互特征融合模块，通过动态调整注意力区域提升密集小目标检测精度。实验结果表明，所提方法在输电线路巡检任务中的综合性能优于当前主流模型，可有效识别输电线路中的小目标部件，有助于发现潜在的安全隐患。更多还原

【Abstract】 Aiming at the problem of low small target detection accuracy caused by complex background interference, large target scale variations, and densely distributed components in transmission line inspection(TLI), this paper proposes a detection transformer(DETR) model named TLI-DETR for TLI tasks. Firstly, multi-scale separable convolutional networks are used to replace the residual blocks at the final layer of each stage in the backbone network. The feature maps output from each stage are extracted and fused as multi-scale features to enhance small target feature extraction capability. Secondly, a channel-spatial fusion cross-attention mechanism is introduced into the multi-scale feature maps to suppress interference from background noise on small target detection. Finally, a hierarchical interactive feature fusion module is designed to improve the detection accuracy of densely distributed small targets by dynamically adjusting attention regions. Experimental results demonstrate that the proposed method outperforms current mainstream models in comprehensive performance for TLI tasks, effectively identifying small target components in transmission lines and facilitating the detection of potential safety hazards.更多还原