Enhancing symbolic image classification through Gaussian copulas and optimized distinguishing points.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Traditional image classification relies heavily on pixel intensity values.
• Symbolic data representations offer a more generalized characterization of pixel intensity patterns.
• Existing methods often use pre-determined distinguishing points, limiting adaptability.

Purpose of the Study:

• To propose a new image classification method utilizing symbolic data (ECDFs and DFDVs).
• To develop a clustering-based approach for selecting distinguishing points to maximize class separability.
• To enhance image feature characterization beyond raw pixel intensities.

Main Methods:

• Utilized empirical cumulative distribution functions (ECDFs) and distribution functions of distribution values (DFDV) as symbolic features.
• Implemented a clustering-based approach to identify optimal distinguishing points for DFDV creation.
• Integrated clustering-based symbolic feature extraction with copula-based modeling.
• Evaluated the method on the MNIST handwritten digits dataset.

Main Results:

• Achieved an average classification accuracy of 68.27% on the MNIST dataset.
• Reached a highest classification accuracy of 95.33% on the MNIST dataset.
• Demonstrated the effectiveness of the proposed symbolic feature extraction and modeling approach.

Conclusions:

• The proposed method offers a competitive and promising alternative for image classification tasks.
• Symbolic data representations, particularly DFDVs derived from optimal distinguishing points, enhance feature characterization.
• The integration of clustering and copula-based modeling shows significant potential for improving classification accuracy.