SFE-DETR: An Enhanced Transformer-Based Face Detector for Small Target Faces in Open Complex Scenes

Area of Science:

• Computer Vision
• Deep Learning

Background:

• Small face detection in complex scenes presents challenges like occlusions and scale variations.
• Existing methods struggle with accuracy and efficiency in dense, degraded environments.

Purpose of the Study:

• To develop an effective and efficient small face detector for open, complex scenes.
• To improve detection accuracy and computational performance simultaneously.

Main Methods:

• Utilized an inverted residual shift convolution and dilated reparameterization backbone.
• Incorporated a multi-head multi-scale self-attention mechanism for feature fusion.
• Introduced a redesigned SFE-FPN with high-resolution layers and a novel feature fusion module.

Main Results:

• SFE-DETR reduced parameters by 28.1% compared to RT-DETR-R18.
• Achieved 94.7% mAP50 and 42.1% AP-s on SCUT-HEAD dataset.
• Attained 86.3% mAP50 on WIDER FACE (Hard) subset.

Conclusions:

• SFE-DETR demonstrates superior detection performance for small faces in challenging conditions.
• The model achieves optimal results for its scale, balancing accuracy and efficiency.