Action registration method and system based on small sample machine learning model

Abstract

The invention discloses an action registration method and system based on a small sample machine learning model. The method comprises the following steps: pre-training a video convolutional network; acquiring a sample data set containing representative actions, extracting features of video samples in the data set by using the video convolutional network, constructing data pairs based on the features and labels of the video samples corresponding to the features, and registering the data pairs in an action library; constructing an error rejection rate and an error acceptance rate based on the data pair, and generating a similarity threshold based on the error rejection rate and the error acceptance rate; and performing feature extraction and discrimination on the real-time video based on the video convolutional network, the action library and the similarity threshold, and outputting a behavior tag.

Description

technical field [0001] The invention relates to the technical field of action recognition, in particular to an action registration method and system based on a small-sample machine learning model. Background technique [0002] Traditional vision technology requires a large amount of labeled data as support. In the case of limited training samples, the generalization effect of deep learning is usually poor. However, in many actual application scenarios, there is often not a large amount of labeled data related to tasks, and the cost of manual labeling is very high, and it is unrealistic to manually complete various types of data labeling. How to endow the model with the ability to learn quickly and recognize new entities with only a small amount of data has become an urgent problem to be solved. [0003] This ability is easy to achieve for humans, but relatively difficult for deep neural networks. The following factors make it difficult to generalize small samples. 1) Deep ...

AI Technical Summary

Problems solved by technology

[0006] In view of this, the present invention provides an action registration method and system based on a small-sample machine learning model. On the basis of the pre-training model, only a small number of video samples are needed for action feature extraction and registration, and subsequent video stream actions can be performed in real time. Recognition, which solves the problems of large samples and poor generalization ability in the traditional neural network model training method for action recognition

Images