Abstract: Course teaching quality is a core content to measure the teaching level of a school, and teaching effect can be directly reflected from the state of students’ attendance. In order to improve students’ class status and promote class discipline, this paper proposes a detection method for students’ classroom behavior based on Faster R-CNN and transfer learning with multi-channel feature fusion. Firstly, images are obtained through the monitoring video of Zhejiang Agriculture and Forestry University and manually annotated, and data augmentation method is used to increase the scale of the images to establish the dataset of common students’ normal classroom behavior. Then, the Inception-ResNet-v2 network based on pre-training is applied for feature extraction, and the target detection framework adopts Faster R-CNN to realize the detection of normal learning, sleeping, lowering head and other student behaviors through transfer learning. Finally, through multi-channel feature fusion method, the shallow features of more detailed information are integrated in deep layers of rich semantic information, so as to gain the improved detection model of the students’ classroom performance. Experimental results show that the mean average precision of the model can reach 76.32%, which is 12.22 percentage points higher than original algorithm, and good detection effect can be achieved. This model has a high accuracy rate for students’ classroom behavior, which indicates that Faster R-CNN with multi-channel feature fusion has a good application prospect in students’ classroom behavior detection, and can provide a new reference for improving classroom teaching quality.

Key words: classroom behavior detection, Faster R-CNN, feature fusion, migration learning