A key frame extraction method for ship surveillance video based on bidirectional GRU and attention mechanism

Abstract

The invention discloses a ship monitoring video key frame extraction method based on a two-way GRU and an attention mechanism, comprising the following steps: step S1: training a two-way GRU and an attention mechanism model by using a large number of seaside ship monitoring video data sets. Step S2, extracting key frames from ship monitoring video by using trained two-way GRU and attention mechanism model, and providing a set of key frames for quickly retrieving ship. By adopting the technical proposal of the invention, the key frames are extracted and applied to the ship monitoring video, a large number of redundant video frames are eliminated, efficient search and browsing of ship events are provided, and the overhead of video storage is saved. At the same time, two-way GRU and attentionmechanism are used to model the relationship between the video frames, and the temporal information is fused into the model, and the information at each time is given different weight, that is, eachtime gives different attention degree, so that the key frame set which is more consistent with human semantics can be obtained.

Description

technical field [0001] The invention relates to fast retrieval based on ship video content and lightweight storage of ship data, in particular to a ship monitoring video key frame extraction method based on a two-way GRU and an attention mechanism. Background technique [0002] Video has become one of the most common sources of visual information. The scale of video data is rapidly expanding. For the videos uploaded to Youtube every day, individual users need more than 100 years to watch all of them. Then automatic tools for analyzing and understanding video content are very important. In particular, automatic video keyframe extraction technology can help users browse video data. A set of video key frames with good effect can succinctly represent the original video, extract important events, and represent the content of the original video with a short and readable key frame summary. With the deployment of seaside surveillance cameras, a large amount of video of ships is g...

AI Technical Summary

Problems solved by technology

[0004] 1. The existing key frame extraction often ignores the connection between video frames. For the video key frame extraction of video semantics, we not only need to use the visual features of the previous video frame, but also need to use the relationship between video frames. The connection in time will largely deviate from the extraction standard of video key frames

[0005] 2. For the extraction of video key frames, attention problems in video frame sequences are often ignored. Adjacent frames may gain and suppress the keyness of the frame. The lack of modeling of this attention mechanism is critical for key frames. Frame extraction has negative effects

Images