Deep Learning and Sensing Technologies in Higher Education: Classroom Behavior Recognition and Evaluation by DA-DBN with LHHO Optimization

Resumen

Student engagement plays a crucial role in facilitating comprehension in the teaching-learning process. Evaluating students' classroom behavior is essential for assessing the effectiveness of teaching. To objectively assess students' attentiveness in class, it is imperative to capture their emotions and cognitive feedback through expressions and behaviors. However, accurately extracting expression features from classroom videos poses a importantcontest. To discourse this, we suggest a hybrid optimization and deep learning technique used for higher education students. Our model involves designing a hybrid spatio-temporal ResNet with Laplacian Harris hawk optimization (LHHO) algorithm for feature extraction and optimization. It enables the extraction of contextual, semantic, and positional features while enhancing the recognition of occluded objects. Additionally, we employ a domain adaptive deep belief network (DA-DBN) for classroom behavior recognition and evaluation, improving the detection accuracy of occluded objects. To validate our approach, we evaluate it against existing classroom behavior recognition models using public datasets such as COCO and PASCAL VOC. The results demonstrate that our proposed algorithm accurately detects various student behaviors, outperforming state-of-the-art models. Specifically, LHHO+DA-DBN model achieves an average accuracy of 96.031% and a recall of 94.335%, surpassing the performance of compared models across different scenarios.