Face Detection With Different Scales Based on Faster R-CNN.

Area of Science:

• Computer Vision
• Deep Learning
• Artificial Intelligence

Background:

• Deep convolutional neural networks achieve high success in face detection.
• Detecting small-scaled faces (less than 15x15 pixels) remains a significant challenge due to feature map shrinking in deep networks.
• Existing scale-invariant methods struggle with extremely small faces.

Purpose of the Study:

• To propose a novel face detection method, the Different Scales Face Detector (DSFD), specifically designed to address the challenge of small-scale face detection.
• To enhance the precision of face detection for small faces without compromising real-time performance.
• To improve upon existing state-of-the-art face detection techniques.

Main Methods:

• Developed a Different Scales Face Detector (DSFD) based on the Faster R-CNN architecture.
• Introduced an efficient multitask region proposal network (RPN) for obtaining human face regions of interest (ROIs) and generating anchors.
• Proposed a parallel-type Fast R-CNN network that assigns proposals to three distinct networks based on their scale, differing in feature map concatenation weights.

Main Results:

• The DSFD method demonstrates improved precision in face detection, particularly for small-scaled faces.
• The proposed network achieves real-time performance comparable to standard Faster R-CNN.
• DSFD shows promising performance on popular benchmarks such as FDDB, AFW, PASCAL faces, and WIDER FACE, outperforming methods like UnitBox, HyperFace, and FastCNN.

Conclusions:

• The DSFD method effectively tackles the challenge of small-scale face detection.
• The integration of multitask learning, feature pyramid, and feature concatenation strategies contributes to enhanced detection accuracy.
• DSFD offers a robust and efficient solution for real-time face detection, especially in scenarios with varying face sizes.