A novel parameter dense three-dimensional convolution residual network method and its application in classroom...

Area of Science:

• Artificial Intelligence
• Computer Vision
• Educational Technology

Background:

• Traditional classroom quality analysis methods are subjective and resource-intensive.
• Computer vision (CV) offers a solution for objective, real-time classroom monitoring.
• There is a need for accurate student expression and behavior recognition systems.

Purpose of the Study:

• To propose a novel Dense 3D Convolutional Residual Network (D3DCNN_ResNet) for analyzing student expressions and behaviors.
• To improve the rationality and accuracy of classroom quality analysis.
• To leverage CV for enhanced teaching strategies via real-time student engagement feedback.

Main Methods:

• Combined Single Shot Multibox Detector (SSD) with an improved D3DCNN_ResNet.
• Utilized 3D convolution in spatial and temporal domains.
• Incorporated residual blocks with dense connections for feature flow and network depth.

Main Results:

• Achieved 97.94% accuracy for expression recognition (CK+ dataset).
• Reached 98.86% accuracy for behavior recognition (KTH dataset).
• Demonstrated efficient model training and improved recognition accuracy.

Conclusions:

• The D3DCNN_ResNet effectively recognizes student expressions and behaviors.
• The network's architecture enhances feature flow and reduces redundancy.
• This technology is suitable for classroom quality analysis and improving teaching strategies.