Block-wise two-dimensional maximum margin criterion for face recognition.

Area of Science:

• Computer Vision
• Machine Learning
• Pattern Recognition

Background:

• Traditional Maximum Margin Criterion (MMC) is effective for feature extraction but limited to vector data.
• Existing methods struggle to capture local characteristics inherent in image data.
• There is a need for dimensionality reduction techniques that can process matrix-formatted image data effectively.

Purpose of the Study:

• To propose a novel two-dimensional generalized framework for Maximum Margin Criterion (MMC) tailored for image data.
• To develop a block-wise approach (B2D-MMC) that leverages local image characteristics for enhanced feature extraction.
• To achieve efficient and effective dimensionality reduction for image datasets.

Main Methods:

• Introduced Block-wise Two-Dimensional Maximum Margin Criterion (B2D-MMC) for matrix data (images).
• Employed a block-wise strategy to exploit local image features within defined blocks.
• Utilized unilateral matrix multiplication for local subspace projections, avoiding iterative processes.
• Sought a closed-form solution for the one-side projection matrix in each block set.

Main Results:

• B2D-MMC successfully extracts local subspace projections from image blocks.
• The method ensures blocks of the same class are close and different classes are far in the projected subspace.
• Experiments on benchmark face databases demonstrated the effectiveness and efficiency of B2D-MMC.
• The proposed method avoids the iterative and alternating procedures common in bilateral projection techniques.

Conclusions:

• B2D-MMC offers an effective and efficient solution for feature extraction and dimensionality reduction in image analysis.
• The block-wise approach overcomes the limitations of traditional MMC for matrix-based data like images.
• The closed-form solution contributes to computational efficiency compared to iterative methods.