Video classification method based on recurrent neural network

Abstract

The invention discloses a video classification method based on a recurrent neural network. On the basis of existing GRU video classification, in the GRU training process, Taylor series are innovatively used for explaining a gating cycle unit, and parameter training of the GRU is assisted by introducing remainder items in the Taylor series in the GRU training process. In the Taylor series, the error is related to the order of the series, and the larger the order is, the smaller the error is, so that the video classification error is reduced by constructing the high-order Taylor series for assisting the GRU training. Meanwhile, the relation between the gating circulation unit and the Taylor series is established, and further, the generalization ability of the GRU is explained through the relation.

Description

technical field [0001] The invention belongs to the technical field of video information mining, and more specifically relates to a video classification method based on a cyclic neural network. Background technique [0002] Video classification refers to classifying the content contained in a given video segment. The categories are usually action (such as making a cake), scene (such as the beach), object (such as a table), etc. Among them, video action is the most popular category. After all, the action itself contains "dynamic" factors, not "static" Images can describe it. [0003] Existing video classification methods mainly include sequential modeling methods based on Recurrent Neural Networks (RNN for short) and sequential modeling methods based on Convolutional Neural Networks (CNN for short). The RNN-based sequence modeling method is to sample the video sequence frame by frame, and use repeated calculation modules to repeatedly calculate, so as to obtain the classifi...

AI Technical Summary

Problems solved by technology

[0004] Compared with CNN-based sequence modeling methods, RNN-based sequence modeling cannot obtain global information, so it is often accompanied by large errors

The CNN-based sequence modeling method cannot classify videos in real time, and must obtain a complete fixed-length video sequence. For example, in a real scene, a surveillance camera needs to judge abnormal events in real time. The model method can only set fixed judgment nodes, and cannot respond to events in real time

[0005] Although other improved video classification methods based on RNN or CNN will improve the accuracy, the improvement of model building is often accompanied by huge consumption of computing resources and time resources, making these complex models unable to be effectively deployed on low-cost devices

[0006] At the same time, existing video classification methods, whether based on RNN, CNN-based sequence modeling methods, or improved methods of these two methods, are often considered to lack interpretability.

Images