Facial expression recognition in videos using hybrid CNN & ConvLSTM.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Video-based facial expression recognition (VFER) is crucial for human-computer interaction.
• Traditional methods like fully-connected LSTM (FC-LSTM) lose spatial information.
• Convolutional LSTM (ConvLSTM) preserves spatial data by integrating LSTM with convolutions.

Purpose of the Study:

• To propose a novel hybrid neural network architecture for VFER.
• To leverage the strengths of 3D-CNN and ConvLSTM for improved spatiotemporal feature extraction.
• To achieve competitive accuracy and efficiency in facial expression recognition.

Main Methods:

• A hybrid architecture combining 3D-CNN and ConvLSTM was developed.
• The model processes video sequences to capture dynamic emotional cues.
• Experiments were conducted on public datasets: SAVEE, CK+, and AFEW.

Main Results:

• The proposed hybrid model achieved competitive accuracy on multiple VFER datasets.
• The architecture demonstrated excellent performance without requiring external emotional data.
• The model is simpler, has fewer parameters, and offers significant speed improvements over state-of-the-art deep learning models.

Conclusions:

• The hybrid 3D-CNN and ConvLSTM model offers an effective solution for VFER.
• The pipeline provides a balance of high recognition accuracy and computational efficiency.
• This approach is suitable for real-time facial expression recognition on resource-limited embedded systems.