Robust LSTM-Autoencoders for Face De-Occlusion in the Wild

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Face recognition is crucial for security but struggles with partial occlusion.
• Existing de-occlusion methods often fail in unconstrained, real-world scenarios.
• Robust face recognition under occlusion remains a significant challenge.

Purpose of the Study:

• To propose a novel Robust LSTM-Autoencoders (RLA) model for effective face de-occlusion.
• To enhance face recognition accuracy for partially occluded faces in real-world applications.
• To develop a method robust to various occlusion types and locations.

Main Methods:

• A Robust LSTM-Autoencoders (RLA) model with two LSTM components: a multi-scale spatial LSTM encoder and a dual-channel LSTM decoder.
• The encoder processes facial patches at various scales for occlusion-robust encoding.
• The decoder reconstructs faces and detects occlusion iteratively, enhanced by an identity-preserving adversarial training scheme.

Main Results:

• The RLA model effectively restores partially occluded faces across diverse datasets (synthetic and real).
• Demonstrated superior performance in removing various facial occlusions compared to existing de-occlusion methods.
• Significantly improved face recognition performance on recovered, partially occluded faces.

Conclusions:

• The proposed RLA model offers a robust solution for face de-occlusion in unconstrained environments.
• RLA enhances the practical applicability of face recognition in surveillance and security.
• The method achieves high accuracy and preserves identity information in recovered faces.