Violent video classification method and system and storage medium

Abstract

The invention provides a violent video classification method and system integrating internal and external knowledge, a storage medium and computer equipment. The method comprises the steps of extracting features of an RGB mode, a motion mode and an audio mode of a multi-label violent video, and establishing multi-mode feature fusion network feature fusion; providing an internal knowledge module for self-adaptively learning multi-scale violent video features, so that richer and more key information can be learned; introducing an external knowledge module for improving the model from high-levelsemantics; constructing a violence-related violence knowledge graph, and proposing an optimization strategy for eliminating redundancy between different labels to obtain a violence correction matrix correction model; introducing a fusion score smoothing strategy to eliminate prediction errors, and building a violent video classification model integrating internal and external knowledge models. According to the violent video classification scheme designed by the invention, the effectiveness and stability of violent video classification are improved.

Description

technical field [0001] The invention belongs to the technical field of video classification, in particular to a violent video classification and recognition method incorporating external knowledge and external knowledge Background technique [0002] The development of technologies such as the Internet has led to an explosive growth of multimedia video content on the Internet. In recent years, with the high popularity of domestic video live broadcasts, many unpredictable violent content often appears. Public safety is an important guarantee for maintaining social stability, and video surveillance is an effective measure to prevent violent crimes. The size of the city and the large number of surveillance situations make it almost impossible to artificially monitor violence. The movie industry produces thousands of movies every year of different genres, and not all movies are suitable for teens and children, especially movies with violent scenes. Regular viewing of violent mov...

AI Technical Summary

Problems solved by technology

[0004] In addition, the current violent video classification mainly relies on violent video features obtained from limited labeled training data, but the features and knowledge learned by this method are limited by the size and distribution of training data.

Unable to effectively integrate external knowledge and labeled data information, lack of network semantic reasoning ability

At present, there are also academic organizations that embed object attributes, label relationship knowledge, and pairwise relationship co-occurrence knowledge, etc., to produce large-scale data sets with external knowledge such as VG and ADE, but it requires a lot of manpower and cost, and the model does not have universal knowledge. ability

Images