Discriminative Fisher Embedding Dictionary Learning Algorithm for Object Recognition.

Area of Science:

• Machine Learning
• Computer Vision
• Signal Processing

Background:

• Discriminative Dictionary Learning (DDL) is vital for classification.
• Current DDL methods often overlook the interplay between dictionary atoms and coding coefficients' interclass and intraclass properties.

Purpose of the Study:

• To propose a novel Discriminative Fisher Embedding Dictionary Learning (DFEDL) algorithm.
• To simultaneously model Fisher embedding for both learned dictionary atoms and coding coefficients.

Main Methods:

• Constructed a discriminative Fisher atom embedding model using the Fisher criterion for atoms.
• Formulated a discriminative Fisher coefficient embedding model on coding coefficients and profiles.
• Ensured the coding coefficient matrix becomes block-diagonal via the Fisher criterion.

Main Results:

• The proposed DFEDL algorithm demonstrated superior performance compared to state-of-the-art methods.
• Achieved significant improvements on both hand-crafted and deep learning-based features.
• The dual Fisher embedding models interactively enhance dictionary and coefficient discriminative capabilities.

Conclusions:

• DFEDL effectively combines interclass and intraclass information for improved classification.
• The simultaneous modeling of atoms and coefficients offers a significant advancement in dictionary learning.
• DFEDL shows promise for various feature types in classification tasks.