A Short Video Retrieval Method Based on Spatio-Temporal Feature Fusion

With the explosive growth of the number of user-generated contents on short video platforms, video retrieval algorithms are widely used in short video copyright protection. In this paper, a short video retrieval method based on spatio-temporal feature fusion is proposed, aiming to solve the copyright protection problem of short video works. Firstly, this method constructs the spatial similarity matrix between short video frames by using semantic features extracted from ResNet50 to represent the content information of short videos. At the same time, the neighboring frame differences are computed to construct the time domain similarity matrix, which represents the dynamic information of the short video. Then, the binary fingerprint of short video content is generated by hashing and dimensionality reduction, and the inter-video similarity is calculated based on the fingerprint similarity. Then, the similarity threshold is adjusted on the training and validation sets to optimize the detection performance. Finally, a lightweight student model is trained by knowledge distillation to achieve efficient operation of video plagiarism detection. Comparative tests prove that this method can significantly improve the accuracy and computational efficiency of short video copyright retrieval. Finally, a lightweight student model is trained by means of knowledge distillation to realize the efficient operation of video plagiarism detection. We apply the model to a real short video copyright retrieval task to verify the effectiveness and practicality of the method.