End2End Occluded Face Recognition by Masking Corrupted Features.

Area of Science:

• Computer Science
• Artificial Intelligence
• Biometrics

Background:

• Deep convolutional neural networks have advanced general face recognition.
• Current models struggle with occluded faces, a common real-world challenge.
• Lack of large-scale occluded face datasets and specific occlusion-handling designs hinder performance.

Purpose of the Study:

• To develop a novel face recognition method robust to occlusions.
• To address the limitations of existing models in handling corrupted facial features due to occlusions.

Main Methods:

• Introduced FROM (Face Recognition with Occlusion Masks), an end-to-end deep neural network.
• Developed a method to dynamically learn masks for cleaning corrupted features within deep convolutional neural networks.
• Constructed massive datasets of occluded face images for effective training.

Main Results:

• FROM demonstrates robustness to occlusions, significantly improving accuracy.
• The method generalizes well to general face recognition tasks.
• Experimental results validated on LFW, Megaface Challenge 1, RMF2, and AR datasets.

Conclusions:

• FROM offers a simple yet powerful solution for occluded face recognition.
• The dynamic masking approach effectively cleans corrupted features.
• The method outperforms existing approaches that rely on external detectors or shallower models.