Deep Cascade Model based Face Recognition: When Deep-layered Learning Meets Small Data.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Sparse Representation Classification (SRC), Nuclear-Norm Matrix Regression (NMR), and Deep Learning (DL) are successful in face recognition (FR).
• Existing methods have limitations: SRC/NMR are one-step, not fully exploiting coding errors; DL requires large data and computation for back-propagation, making it infeasible for small datasets.
• Developing efficient algorithms for small-scale data is crucial for advancing FR.

Purpose of the Study:

• To propose an end-to-end deep cascade model (DCM) for corrupted face recognition, specifically adapted for small-scale datasets.
• To overcome the limitations of existing SRC, NMR, and DL methods in resource-constrained scenarios.
• To demonstrate a new approach to deep-layered learning that does not rely on convolutional neural networks or back-propagation.

Main Methods:

• An end-to-end deep cascade model (DCM) is proposed, integrating hierarchical learning, nonlinear transformation, and a multi-layer structure.
• A multi-level pyramid structure is incorporated for enhanced local feature representation.
• Softmax vector coding of errors with class discrimination is introduced for nonlinear transformation in layer-wise learning.

Main Results:

• The proposed DCM achieves superior performance on small-scale benchmark face recognition datasets compared to state-of-the-art methods.
• The model demonstrates effectiveness in corrupted face recognition scenarios.
• Experiments validate the model's ability to learn powerful representations without extensive data or back-propagation.

Conclusions:

• The developed DCM offers an efficient and effective solution for face recognition with limited data.
• The research consolidates the perspective that deep-layered learning can be achieved without traditional convolutional neural networks and back-propagation.
• The proposed framework allows for easy integration of existing representation methods, enhancing its versatility.