Research On Student Behavior Recognition Based On Semi-supervised Transfer Learning

With the development of science and technology,computer vision technology has gradually stepped into everyone's work and study.In the field of computer vision,using machine learning to recognize human features is one of the hot and difficult research topics.Behavior recognition technology is common in daily life,and is becoming more and more popular in life and work.For example,it has a very long-term development prospect in human-computer interaction,automatic driving,automatic monitoring of industry and agriculture,etc.In behavior recognition,migration and a semi-supervised learning is more effective under the field of two kinds of machine learning method,can use have data to the computer in the field of autonomous learning with areas and across similar behavior,this solves the less training samples in the video and image caused by the problem of low efficiency of learning.And with the increasing popularity of education,the country also attaches more importance to education,and the number of college students is gradually increasing.Therefore,in order to improve the efficiency of student campus security and behavior recognition,this paper proposes a semi-supervised transfer learning algorithm for student behavior recognition in the same field.In this paper,the transfer learning method is used to improve the data redundancy and the few samples in the training data.Secondly,in the process of extraction of characteristic value,in view of the convolution neural network to extract the feature need to appear in the process of the fixed input image problems setting up the sample size and the size of the convolution kernel,this paper USES a space technology to the input data of the pyramid of image scale,make convolution layer can better carry on the training.Finally,a KNN model is constructed to optimize the training process of convolutional neural network and improve the accuracy and efficiency of recognition.In the experimental stage,this paper designed seven groups of comparative experiments.Firstly,six kinds of student behavior images collected were used to verify the transfer learning of UCF data set and ASD data set.The effect was good and the feasibility of transfer learning was confirmed.Secondly,six types of student behaviors in UCF101 dataset were used to conduct semi-supervised transfer learning experiments on ASD dataset and Weiziman dataset respectively,and the accuracy was compared with other algorithms.Experimental results demonstrate the effectiveness of the proposed algorithm.