2D Feature Selection by Sparse Matrix Regression.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Traditional methods for matrix data in image processing often convert matrices to vectors, losing spatial information and increasing dimensionality.
• Direct feature selection for 2D matrix data remains an underexplored but critical issue.

Purpose of the Study:

• To propose a novel algorithm, sparse matrix regression (SMR), for direct feature selection on matrix-form data.
• To address the limitations of vector-based approaches in handling matrix data by preserving spatial information.

Main Methods:

• Developed the sparse matrix regression (SMR) algorithm, which accepts matrix data as direct input.
• Incorporated sparse constraints on regression coefficients for effective feature selection.
• Proposed an optimization method with proven convergence for the SMR algorithm.

Main Results:

• Demonstrated that the number of regression vectors acts as a crucial parameter for balancing learning capacity and generalization.
• Empirically validated SMR's effectiveness against several vector-based methods on benchmark datasets.
• Showcased SMR's successful application in scene classification tasks.

Conclusions:

• Sparse matrix regression (SMR) offers a powerful and effective solution for direct feature selection on matrix data.
• SMR outperforms traditional vector-based methods by preserving crucial spatial information inherent in matrix data.
• The proposed method shows significant promise for various image processing and computer vision applications.