Face Detection Algorithm Based on Double-Channel CNN with Occlusion Perceptron.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Deep Learning

Background:

• Face detection accuracy is significantly reduced by complex occlusion conditions.
• Existing methods struggle with partially or fully occluded faces.
• Robust face detection under occlusion remains a critical challenge in computer vision.

Purpose of the Study:

• To propose a novel neural network model for accurate face detection under complex occlusion.
• To enhance the feature extraction capabilities for occluded facial regions.
• To improve the overall performance and speed of face detection systems.

Main Methods:

• A double-channel occlusion perceptron neural network model was developed, integrating an area occlusion judgment unit into the VGG16 network.
• Feature extraction for unoccluded and less occluded facial regions was performed using the perceptual neural network.
• Transfer learning was employed for pretraining convolutional layer parameters to mitigate overfitting.
• Optimized residual network was used for whole-face feature extraction, followed by weighted fusion with occlusion perceptron features.

Main Results:

• The proposed method demonstrated higher detection accuracy compared to existing approaches on the AR and MAFA datasets.
• The model achieved faster detection speeds, indicating computational efficiency.
• Experimental results validate the effectiveness of the occlusion perceptron and feature fusion strategy.

Conclusions:

• The developed double-channel occlusion perceptron neural network effectively addresses the challenge of low accuracy in face detection under complex occlusion.
• The fusion of features from the occlusion perceptron and residual network significantly improves detection performance.
• The method offers a promising solution for real-world face detection applications with occluded faces.