Abstract: The intelligent detection of the edge of power instrument equipment is a necessary link in the construction of digital substations. When the power instrument is detected by the mobile visual equipment, it is difficult for the edge computing of the equipment to realize the rapid detection of small scale and high likelihood target images in complex environments. Therefore, a target detection of the power instrument images based on the lightweight EF-YOLOv4 network is proposed in this paper. By improving the backbone feature network of the model, this method uses the deep separable convolution to extract the multi-attribute features of the instruments, reduces the computational complexity of the model and improves the detection speed. By improving the feature fusion structure and adding a shallow feature layer with high resolution, color, texture and other instrument information, the model is improved in its attention to the small-scale instrument targets. The method of fast library for approximate nearest neighbors (FLANN) is integrated to detect the instrument targets through the fine-grained features of the unit symbols. The transfer learning parameter sharing mechanism is used to adjust the weight of the model and the model will quickly adapt to the small sample data set of the power instruments. Finally, the power instrument image test set is constructed to verify the model. The experimental results show that compared with the traditional target detection methods, the proposed method maintains high accuracy and speed for target detection of the multi-scale and fine-grained instrument equipment images such as power meters and voltmeters. It will provide feasible technical solution and reference for the visualization, informatization and intelligentization of the power instruments.