Substation meter recognition is very significant to status monitoring of instrument equipment in power system. However, meters in substation are various, small and dense distribution. Therefore, in the process of recognizing meter types based on artificial intelligence, there are many problems, such as low pixel, few key features and a great deal of confusion. How to overcome these problems and improve accuracy is critical. Moreover, Algorithm deployment on edge device is also in trouble. In this paper, a lightweight meter recognition Yolov5-based algorithm is proposed. Firstly, Transformer block is selected to enhance the ability of global information extraction and establish long-term dependence. Secondly, convolution block attention mechanism (CBAM) is supplied to strength the power of distinguishing. It can integrate underlying information effectively and be conducive to small object detection. In addition, Mish and E-IoU are used to improve convergence rate, which ensured accuracy and improved model detection speed. Finally, for edge deployment, Ghost module is fused to reduce model parameters and simplify network structure. Experiment results show that the improved method has satisfactory speed and surprising accuracy in the recognition of pointer-type meter in complex substation.