Normal operation of high-voltage disconnect switches is a prerequisite for the stable operation of power systems. To correctly identify the breaking and closing states of disconnect switches, a lightweight improved YOLOv5s target detection algorithm is proposed. First, the anchor frame parameters are reacquired using a quadratic optimized K-means++ clustering algorithm for the disconnecting switch dataset. Then, the loss function in the model is replaced from CIOU to EIOU with stronger convergence performance to accelerate the convergence speed of model training. Finally, a CBAM attention module is added to the last layer of the model backbone feature extraction network to strengthen the model feature extraction capability. The improved model is lightened using the channel sparsification pruning method to reduce the model size and arithmetic power consumption. The experimental results show that the average accuracy of the improved model reaches 97.4% and the model's size is 3.92 MB after the light weighting process, making the model easier to deploy to mobile devices for real-time detection.