Deep Hybrid YOLOv5-Capsule Network for Group Facial Expression Recognition

Abstract:
In recent years, facial expression recognition has become a rapidly growing field of research. Modern communication relies on speech, text, and facial expression recognition, making it essential to understand emotions and implicit expressions. However, variations in lighting conditions, image quality, and geometry further compound the problem. To address the issue presented in this paper, the proposed YDn-CNN model integrates YOLOv5 for real-time object detection and DenseNet for deep feature extraction. Furthermore, YOLOv5 provides faster detection of facial regions within the group and improved accuracy. Moreover, the Adaptive Gaussian-Contrasting Wavelet Filter (AGC-WF) model can be used to preprocess images obtained from the dataset, enhancing facial features in video frames. Then, the Entity Scalar Slice Window Segmentation (ESSWS) technique is used to isolate individual facial regions from each frame. Additionally, the Entity Scalar Particle Swarm Intelligence (ESPSI) technique is employed for optimal feature selection. Finally, the proposed YDn-CNN model combines YOLOv5 for real-time object detection and DenseNet for deep feature extraction. The proposed system demonstrates higher precision, recall, and F1-score compared to existing facial expression recognition models. Furthermore, the proposed approach ensures improved expression detection accuracy of up to 96%.