Scene-Level Recognition of Classroom Operational States Using VGG16 Transfer Learning

Abstract

Automatic recognition of classroom operational states can be useful for classroom availability, room-use monitoring, and academic administration. This paper investigates whether three general classroom states can be recognized using a deep learning-based computer vision approach. The three states are lecture in progress, students present but the teacher absent, and empty classroom. The dataset was collected locally and contains about 1,200 images, with nearly 400 images in each class. The images were preprocessed and then used to train a VGG16 convolutional neural network. The network was initialized with ImageNet weights and adapted for three output classes. The main experiment used an 80:20 training and testing split. On 240 test images, the model obtained an accuracy of 97.08%. The F1-score was 0.96 for lecture in progress, 0.96 for students present without the teacher, and 0.99 for the empty classroom. Most errors occurred between the two occupied classroom states. The empty classroom was identified more accurately. A small web application was also developed to demonstrate how a classroom image can be submitted to the model and a state prediction returned. The results are promising for the dataset, but they should be considered only as a proof of concept. The images were collected from more than one classroom, but no independent external classroom dataset was used for validation. Therefore, the study gives an initial basis for future work on classroom availability, room-use monitoring, and simple operational monitoring