Efficient classification of short videos is of a great challenge in industry due to their large amounts and diverse semantics. In this paper, we present a real-time multi-label classification system to attain it. Specifically, a frame-level preprocessing strategy is first proposed to efficiently decode the videos for useful information. Then an image-based model is developed to achieve the final video-level classification. For the multi-label classification case, three modules are proposed to enhance the model's classification capability, relieve the label-imbalance issue and exploit the label correlation respectively. Our system achieves an accuracy of 86% on both private test sets of the challenge ``Short Video Real-time Classification'' in AI Challenger 2018, and on its open validation set it achieves an accuracy of 89.2% with a speed of 28ms per video on a workstation with an Intel Xeon CPU E5-2650 and a NVIDIA GPU TITAN Xp.