Abstract: It is difficult for mobile edge devices to quickly detect high-similarity targets in complex environments during substation power meter inspections. To address this, we proposed a lightweight YOLOX-based method for power meter image detection. First, we built the YOLOX detection network and designed a depthwise separable convolutional backbone feature extraction structure and a multi-scale feature fusion structure based on parameter reorganization to compress model parameters and improve inference speed. Secondly, a three-dimensional attention mechanism, SimAM, was embedded in the feature fusion layer. This mechanism learns the energy distribution of features and weights the target areas to enhance meter detection in complex environments. Additionally, to address specific issues in power meter detection, we designed a transformer structure based on pyramid pooling feature encoding, focusing on refining local features and capturing long-distance features to mine high-semantic information, thereby improving the detection accuracy of power meters with different shapes. Finally, we validated the algorithm by constructing a dataset of broken, blurred, and normal power pointer meters. The experimental results show that the improved model increases the mean precision from 75.49% to 85.93% and the detection speed from 36 Frame/s to 45 Frame/s compared to the original model. On the mobile hardware Jetson NX, the inference speed reaches 17.6 Frame/s. Compared to other lightweight models, this model has significant advantages in detection accuracy and speed, providing a feasible technical solution for the visualization, informatization, and intelligence of power meters.