Video Segment Retrieval via Scaled-Dot Product Attention

The patent applies preliminary action by pre-training deep neural network algorithms before actual video retrieval operations. The system extracts video features and query features using pre-trained models, and pre-computes attention mechanisms to integrate video-query relations. This preparation work is done in advance, allowing the actual retrieval process to execute quickly without performing heavy computations during real-time queries, thus resolving the contradiction between high retrieval performance and low time consumption.

The patent uses copying by creating feature representations (video features and query features) that capture the essential semantics of the video and query sentences. Instead of processing raw video data repeatedly, the system extracts and stores compressed feature vectors that can be efficiently compared and matched. This copying of essential information into a compact form reduces the computational resources needed for subsequent retrieval operations while maintaining high retrieval performance.

The patent applies parameter changes by transforming the retrieval problem from operating on raw video frames to operating on extracted feature vectors. The system changes the parameter space from pixel-level or frame-level data to semantic feature representations. This parameter transformation allows the system to achieve both high execution speed (by working with compact feature vectors rather than full video data) and high retrieval performance (by using semantically meaningful features captured by deep neural networks).