This paper addresses the urgent need for high-precision and high-efficiency visual perception technologies in power equipment operation and maintenance under the background of rapid development of smart grids. It points out the performance limitations of the existing real-time target detection framework RT-DETR when handling small targets, dense targets, and complex backgrounds in power inspection scenarios. To overcome this bottleneck, this study proposes an improved backbone network model, DETR-EVA, based on an efficient visual attention mechanism (EVA). This model innovatively designs an attention computation structure with linear complexity by deeply integrating the EVA mechanism with the C2f module in the RT-DETR backbone network, and combines local detail perception and global dependency modeling capabilities. Its core lies in the introduction of a gated fusion mechanism, which significantly enhances the model’s ability to model long-distance contextual relationships and the adaptive adjustment efficiency of feature weights while retaining the advantages of multi-branch feature extraction and fusion of the C2f module. Experiments were conducted on an inspection image dataset containing typical power equipment targets. The results show that compared with the original RT-DETR model, DETR-EVA improves the overall accuracy index mAP50-95 by 2.5%, reduces computational complexity by 14%, and reduces the number of model parameters by 27%. This effectively verifies that the proposed method can significantly improve the detection accuracy of small targets and complex scenes while maintaining real-time detection speed, providing a better visual solution for intelligent operation and maintenance of power equipment.