ELT-RTDETR:轻量化变电站缺陷检测的Transformer增强模型

ELT-RTDETR:lightweight Transformer-enhanced model for substation defect detection

摘要:

为解决变电站缺陷检测任务中存在的实时性不足、计算资源消耗高,以及实际环境下检测精度低等问题,提出了一种轻量级目标检测模型ELT-RTDETR。首先,采用EfficientFormerV2作为主干网络,结合局部卷积与轻量化Transformer设计,显著降低模型参数量与计算开销。其次,提出轻量化多尺度特征金字塔网络LMSFPN,通过多尺度深度卷积、加权融合与高效上采样策略,增强多尺度缺陷特征的表达能力,同时减少冗余计算。最后,引入基于TSSA注意力机制的TSSACFE模块,通过局部统计建模与低维投影优化特征交互,有效提升微小缺陷的检测鲁棒性。结果显示,在自建的变电站设备缺陷数据集上,ELT-RTDETR的检测精度达82.1%,较传统RT-DETR提升7.3%,同时模型计算量和参数量分别降低63.2%与50.7%。消融实验与主流算法对比结果表明,该模型在精度、轻量化,以及推理效率上均优于YOLO系列与现有RT-DETR变体,尤其在表计外壳破损和硅胶桶变色等任务中表现突出。本研究为变电站环境下的实时缺陷检测提供了高效解决方案,具备显著的工程应用潜力。

Abstract:

To address the problems of insufficient real-time performance, high computational resource consumption, and low detection accuracy in practical environments for substation defect detection tasks, this paper proposed a lightweight object detection model, namely ELT-RTDETR. First, EfficientFormerV2 was adopted as the backbone network, combining local convolutions with a lightweight Transformer design to significantly reduce the number of model parameters and computational overhead. Second, a lightweight multi-scale feature pyramid network (LMSFPN) was proposed to enhance the expression capability of multi-scale defect features through multi-scale depth-wise convolutions, weighted fusion, and efficient upsampling strategies, while reducing redundant computations. Finally, a token statistics self-attention cross-feature enhancement (TSSACFE) module based on the token statistics self-attention (TSSA) mechanism was introduced. This module optimized feature interaction through local statistical modeling and low-dimensional projection, effectively improving the detection robustness of small defects. Results show that on a self-built substation equipment defect dataset, the detection accuracy of ELT-RTDETR reaches 82.1%, which is 7.3% higher than that of the traditional RT-DETR. Meanwhile, the model calculation volume and parameter count are reduced by 63.2% and 50.7%, respectively. Ablation experiments and comparisons with mainstream algorithms demonstrate that the proposed model outperforms the YOLO series and existing RT-DETR variants in terms of accuracy, light weight, and inference efficiency, especially in tasks such as meter shell damage and silica gel canister discoloration. This study provides an efficient solution for real-time defect detection in substation environments, possessing significant potential for engineering applications.