This research proposes a video content understanding method based on deep learning to effectively understand video content. First, use ResNet to extract global features and Places365-CNNs migration learning to extract deep scene features. Then, scene vectors are generated by multi-layer perceptron and used as the input of LSTM network to encode and decode video images and description statements. Finally, by pre training on the MSCOCO dataset, accurate description statements are generated for video key frames. The experimental results show that with the increase of data volume, the performance of the model is significantly improved, especially in CIDER-D evaluation, which is superior to other models. The conclusion shows that the proposed model shows high accuracy and performance improvement in describing video content.