Graph Regularized Auto-Encoders for Image Representation.

Area of Science:

• Computer Vision
• Machine Learning
• Manifold Learning

Background:

• Image representation is crucial for computer vision tasks like clustering and classification.
• Learning low-dimensional representations that preserve original image information is highly valuable.
• Manifold learning techniques capture intrinsic low-dimensional structures in high-dimensional data.

Purpose of the Study:

• To propose a novel deep nonlinear mapping algorithm for image representation learning.
• To develop a method that preserves local connectivity from the original image space to the representation space.
• To enhance the capacity of deep learning models for complex data modeling and fast inference.

Main Methods:

• A graph regularized auto-encoder (GAE) algorithm is proposed, integrating deep architectures with graph regularization.
• Stacked auto-encoders are utilized to provide an explicit encoding model for efficient inference.
• Graph regularization is applied to preserve local connectivity and penalize the Jacobian matrix norm of the encoder.

Main Results:

• Theoretical analysis reveals the mechanism of the graph regularizer in preserving local properties.
• The method effectively preserves local connectivity in the representation space.
• Experimental results demonstrate GAE's effectiveness in image clustering and classification tasks.

Conclusions:

• The proposed graph regularized auto-encoder (GAE) is an effective method for deep representation learning.
• GAE offers a superior solution compared to existing auto-encoder variants and local invariant methods.
• The theoretical analysis provides insights into the advantages of GAE for image representation.