Tensor Nuclear Norm-Based Multi-Channel Atomic Representation for Robust Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Existing representation-based classification (RC) methods for face recognition often overlook 3D structural correlations in multi-channel color images.
• These methods, typically analyzing grayscale data, struggle with real-world noise like occlusions and corruption, leading to performance degradation.

Purpose of the Study:

• To propose a novel Tensor Nuclear Norm based Robust Multi-channel Atomic Representation (TNN-RMAR) framework for robust color face recognition.
• To address the limitations of existing RC methods in handling multi-channel data and complex noise.

Main Methods:

• A 3D color image-tensor-based error model is proposed to exploit the full 3D structural information of color error images.
• The tensor nuclear norm is utilized to leverage the low-rank property of the 3-order tensor representation of color error images.
• A multi-channel atomic norm (MAN) regularization is devised for representation coefficients to capture inter-channel correlations, alongside a tube-wise loss function and an ADMM-based algorithm.

Main Results:

• The TNN-RMAR framework effectively utilizes 3D structural information across color channels.
• The proposed method demonstrates robustness against various noise types common in real-world face images.
• Experimental results on benchmark databases validate the framework's effectiveness and robustness for color face recognition.

Conclusions:

• The TNN-RMAR framework offers a significant advancement in robust color face recognition by integrating multi-channel information and noise resilience.
• The developed 3D error model and MAN regularization provide a powerful tool for analyzing complex image data.
• The framework serves as a versatile platform for developing new robust multi-channel RC methods.