Color Occlusion Face Recognition Method Based on Quaternion Non-Convex Sparse Constraint Mechanism

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Color face recognition is advancing, with convolutional neural networks showing high performance.
• Deep learning models require large datasets and significant computational resources, hindering accessibility.
• Existing lightweight models like PCANet need improvement in accuracy for occluded color faces.

Purpose of the Study:

• To develop a lightweight and accurate color face recognition method.
• To improve recognition performance specifically under occlusion conditions.
• To address the trade-off between model complexity and recognition accuracy.

Main Methods:

• Proposed a novel color occlusion face recognition method utilizing a quaternion non-convex sparse constraint mechanism.
• Constructed a quaternion non-convex sparse principal component analysis network model with Lp regularization for strong sparsity.
• Employed fixed point iteration and coordinate descent methods to solve the resulting non-convex optimization problem.

Main Results:

• Successfully developed a lightweight model for color face recognition.
• Demonstrated improved recognition accuracy for occluded faces compared to existing methods.
• Validated the method's effectiveness on multiple benchmark datasets: Georgia Tech, Color FERET, AR, and LFW-A Color.

Conclusions:

• The proposed quaternion non-convex sparse method offers an effective solution for lightweight color face recognition, especially under occlusion.
• The study highlights the potential of non-convex sparse constraints in enhancing deep learning models for face recognition tasks.
• This research contributes to more efficient and accurate face recognition systems applicable in resource-constrained environments.