Wasserstein CNN: Learning Invariant Features for NIR-VIS Face Recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Biometrics

Background:

• Heterogeneous face recognition (HFR) faces challenges due to significant intra-class variations and limited cross-modality training data.
• Matching facial images across different sensing modalities, like near-infrared (NIR) and visual (VIS), is crucial for security and forensics.

Purpose of the Study:

• To propose a novel Wasserstein convolutional neural network (WCNN) for effective NIR-VIS face recognition.
• To learn modality-invariant features that bridge the gap between heterogeneous face image distributions.
• To address overfitting issues in HFR with limited training data.

Main Methods:

• A WCNN architecture is employed, with low-level layers trained on VIS data and high-level layers for modality-specific (NIR, VIS) and shared invariant feature learning.
• Wasserstein distance is utilized in the shared layer to minimize dissimilarity between NIR and VIS feature distributions.
• A correlation prior with a low-rank constraint is introduced to regularize fully-connected layers and mitigate overfitting on small datasets.

Main Results:

• The WCNN effectively learns modality-invariant deep feature representations for heterogeneous face images.
• Experimental results on three challenging NIR-VIS face recognition databases demonstrate superior performance compared to existing state-of-the-art methods.
• The proposed method achieves efficient computation for heterogeneous data during the testing phase.

Conclusions:

• The WCNN approach significantly enhances heterogeneous face recognition accuracy by learning robust invariant features.
• The integration of Wasserstein distance and correlation prior offers an effective solution for cross-modality face matching.
• This method shows strong potential for real-world applications in security and surveillance.