Dense pedestrian face detection in complex environments.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Dense crowd face detection in complex environments remains a significant challenge.
• Existing models often struggle with occlusions, varying scales, and cluttered backgrounds.

Purpose of the Study:

• To propose a novel face detection model, Deep and Compact Face Detection (DCFD), for improved performance in dense crowd scenarios.
• To enhance accuracy and efficiency in complex environmental conditions.

Main Methods:

• Utilized an improved lightweight EfficientNetV2 network as the backbone, replacing the original RetinaFace backbone.
• Integrated a large kernel attention mechanism and an improved efficient channel attention (ECA) mechanism.
• Employed an improved neural architecture search feature pyramid network (NAS-FPN) for feature fusion.
• Implemented a focus loss function to balance positive and negative sample training.

Main Results:

• The DCFD algorithm demonstrated efficient and accurate face detection performance across diverse environments.
• Significant improvements in face detection accuracy were observed in various complex scenes.
• The model effectively addresses challenges posed by dense crowds and intricate settings.

Conclusions:

• The proposed DCFD model offers a feasible and effective solution for dense crowd face detection.
• This work provides a strong foundation for advancing the accuracy of face detection models in practical applications.