Effective Face Detector Based on YOLOv5 and Superresolution Reconstruction.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Face detection is crucial for public security systems but struggles with small, dense faces in complex scenes due to image quality and lighting.
• Existing methods face limitations in accurately detecting faces under challenging conditions such as low resolution or blur.

Purpose of the Study:

• To propose a robust face detection model, SR-YOLOv5, specifically designed to overcome the limitations of detecting dense and small faces in real-world scenarios.
• To enhance the performance of face detection in terms of accuracy and speed by optimizing YOLOv5.

Main Methods:

• Optimized the backbone and loss function of YOLOv5 for improved mean average precision (mAP) and speed.
• Integrated image super-resolution technology into the detection head to enhance performance in low-resolution or blurred scenes.
• Trained and tested the SR-YOLOv5 model on a diverse face dataset, comparing it against established algorithms.

Main Results:

• SR-YOLOv5 demonstrated superior detection precision compared to leading algorithms like MTCNN, CMS-RCNN, HR, S3FD, and TinaFace, with improvements of 0.7%, 0.6%, and 2.9% respectively.
• The model effectively handles challenging face detection targets in complex environments.
• Achieved better performance in terms of mAP and speed.

Conclusions:

• SR-YOLOv5 offers a significant advancement in face detection technology, particularly for difficult cases involving small and dense faces.
• The integration of super-resolution enhances robustness in adverse imaging conditions.
• The proposed model contributes to more effective security monitoring and face analysis systems.