Research On Semi-supervised Human Action Recognition Based On Convolutional Neural Network

With the rapid development of the Internet and video capture technology,a large amount of video data is generated online,and how to use intelligent technology to recognize human actions in videos has become an urgent need.Human action recognition has very promising applications in many aspects,many researchers have conducted thoroughly studies and explorations on it,and achieved remarkable research results.Currently,deep convolutional neural networks based on supervised learning have a good application effect on human action recognition tasks,but the recognition performance of such methods depends to a certain extent on the quantity and quality of labeled data,and it is difficult for such methods to show their strong performance when only a small amount of labeled data is available.Compared with supervised learning methods,semi-supervised learning methods can better meet the needs of real-world tasks,which can exploit not only the strongly supervised information of labeled data,but also the unlabeled data to mine more task-relevant and valuable information.In order to decrease the use of label data in human action recognition tasks and reduce the cost of data labeling,this paper studies semi-supervised human action recognition.The main research contents are as follows:(1)Semi-supervised human action recognition based on adaptive correlation learningAiming at the feature modeling problem,considering that labeled samples and unlabeled samples have local feature correlation in the feature space,this paper proposes an adaptive correlation learning module to mine the correlation information between samples,and then uses the obtained correlation information and shallow graph convolution to aggregate the features of video samples in the local neighborhood,generates more expressive features for video samples,and improves the feature discrimination of human action recognition.Experimental results on three publicly available human action recognition datasets,HMDB51,UCF101,and Something-Something V2,show that the performance of the algorithm proposed in this paper outperforms other algorithms on all three datasets.(2)Semi-supervised human action recognition based on contrastive learningAiming at the problem that the performance of supervised deep learning methods is limited in the scenario of a small amount of labeled data,considering that the original samples corresponding to the inputs on different pathways are the same in the two-pathway contrastive learning,and to ensure that the similarity between samples on the two pathways is as consistent as possible,a similarity contrastive learning method is proposed based on the temporal contrastive learning framework,to model the similarity between samples on different pathways using the outputs corresponding to video samples.The experimental results on the publicly available human action recognition dataset Mini-Something-V2 demonstrate that the algorithm proposed in this paper achieves better recognition results with less labeled data.