FERDCNN: an efficient method for facial expression recognition through deep convolutional neural networks.

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Facial expression recognition (FER) is crucial for numerous real-life applications and remains a challenging computer vision problem.
• Deep learning, particularly deep convolutional neural networks (DCNNs), offers promising solutions for complex tasks like FER.

Purpose of the Study:

• To introduce an efficient method for facial expression recognition (FERDCNN) utilizing pre-trained DCNN models.
• To evaluate the performance of FERDCNN on standard datasets (CK+ and JAFFE) for recognizing fundamental emotions.

Main Methods:

• Image pre-processing included face detection, resizing, gamma correction, and histogram equalization.
• Deep features were extracted using five pre-trained DCNN models: AlexNet, GoogleNet, ResNet-18, ResNet-50, and ResNet-101.
• Transfer learning and Support Vector Machine (SVM) were employed for feature classification.

Main Results:

• The FERDCNN method achieved high accuracy rates of 99.0% on the CK+ database and 95.16% on the JAFFE dataset.
• AlexNet performed best as a feature extractor, while SVM excelled as a classifier.
• The combination of AlexNet and SVM yielded the highest recognition accuracies.

Conclusions:

• The proposed FERDCNN method, particularly the AlexNet-SVM combination, is highly effective for facial expression recognition.
• The study validates the efficiency of DCNNs and SVMs in accurately classifying human emotions from facial images.
• This research contributes to advancing automated FER systems for practical applications.