In this paper, research work in intelligent detection and recognition for high-voltage (HV) switch cabinet was introduced with presenting a highly efficient identification algorithm based on deep learning with an improved YOLOv2 convolutional neural network (CNN). Since HV cabinets with many types of switches and their dense distribution, the pooling layer is improved to a space pyramid pooling layer in YOLOv2 network, and a 19×19 mesh partitioning strategy is designed which is more consistent with the recognition of switch state. The proposed algorithm is an end-to-end identification method which can directly implement multiple tasks including switch positioning, recognition, and state determination. Extensive evaluations on real datasets obtained from a substation with 1500 images show that the proposed method achieves the leading performance compared to other existing methods in terms of efficiency and accuracy.