Research On Teachers’ Teaching Pose Recognition Algorithm Based On Convolutional Neural Network

Teachers’ teaching behavior not only guides students’ attention,but also promotes students’ learning performance.This shows the importance of analyzing teachers’ behavior to education.Therefore,we try to apply deep learning technology to the classroom scene to identify teachers’ teaching pose here,so as to analyze the teachers’ teaching behavior.Classroom is the main place where teachers teach.In order to analyze teachers’ teaching behavior,a semantic segmentation algorithm is used to separate the teachers from the complex background of the classroom,and then based on teachers’ attention and gestures,a classification algorithm is used to classify their pose.In response to the lack of a typical classroom scene dataset,classroom dataset for classroom scenarios is created here,which includes classroom-A dataset with pixel-level segmentation labels and classroom-B dataset with classification labels.For the segmentation algorithm,we firstly add the ASPP structure and global pooling layer to the basic network,to capture image features at multiple scales,and integrate global context information.Secondly,the Encoder-Decoder structure is used to fuse image information from shallow network,so that the segmentation network can gradually recover location information.Finally,an auxiliary loss function is added to the original loss function to optimize learning process of the network.As for the classification algorithm,it is mainly based on the fine-tuned advanced classification model and uses the mixup method for data augmentation.Teachers’ pose recognition algorithm is an exploration of the analysis of teachers’ teaching behavior.It promotes the improvement of teachers’ teaching ability through the use of technical analysis.Compared with existing algorithms,the biggest feature is that it can adapt to complex and changing background.The classroom scene segmentation algorithm achieved an optimal MIo U value of 84.95% on the created classroom-A dataset.The average MIo U of five-fold cross-validation exceeded Deep Lab V3+ 1.9%,and FLOPs decreased significantly.The teachers’ pose classification algorithm achieved an average accuracy of 87.83% on the classroom-B dataset,and compared with the average of the three basic networks,the accuracy increased by 1.17%.