The detection of electricity meter data in the distribution box of a power system provides important data support for power management and safe operation. Traditional manual methods for reading electricity meter data are inefficient and prone to errors, while existing deep learning methods are limited in model application due to large model parameter sizes. To address these issues, a lightweight and robust real-time electricity meter detection method was proposed. The parameter size of the model was reduced by reducing the number of layers and channels in the feature extraction network, and the lightweight of the network was achieved. While reducing the network parameter size, the global context and local multi-scale context were introduced to ensure the network feature representation and fitting capabilities. The global context focused on the position of electricity meter data in the meter box, while the local multi-scale context adapted to different sizes of meter data. Experimental results show that the proposed network achieves higher accuracy and faster detection speed than other detection methods, even with smaller parameter sizes.