The switch state recognition is essential for modern power systems, and traditional switch state recognition methods cannot effectively solve the problem of multiple switch target interference. In order to solve the problem, a switch state recognition method based on improved deep learning was proposed. Firstly, a spatially weighted pooling strategy was employed to improve traditional convolutional neural networks (CNNs). Secondly, the model was trained on the training database by using the improved CNNs. Thirdly, the trained model was used to detect the candidate positions of insulators and switches, and then the exactly locations of insulators and switches were extracted via non-maximum suppression algorithm and line fitting method. Finally, the on/off state of switches was recognized by calculating length-width ratio of switch regions and connectivity between switch region and insulator regions. The experiment results show that the proposed method can accurately localize the insulators, switches and significantly improve the precision of recognizing switch state.