Improving class separability using extended pixel planes: a comparative study

Area of Science:

• Computer Vision
• Image Processing
• Machine Learning

Background:

• Feature engineering is crucial for pattern recognition.
• Developing effective feature representations is challenging, especially for novel applications.

Purpose of the Study:

• To investigate class separability using extended representations of pixel planes (EPP).
• To evaluate the impact of different EPP generation methods (scale pyramid, subband pyramid, image transforms) on feature space separability.
• To assess the effectiveness of EPP with suboptimal feature libraries.

Main Methods:

• Generated EPP using scale pyramid, subband pyramid, and various image transforms (Chebyshev, Fourier, wavelets, gradient, Laplacian).
• Explored combinations of these image transforms.
• Evaluated EPP performance with feature libraries containing only textural or Haralick features.

Main Results:

• All three EPP types (scale pyramid, subband pyramid, image transforms) demonstrated improved class separation.
• Transform-based EPP significantly enhanced separability, outperforming scale and subband pyramids, particularly with suboptimal feature libraries.
• EPP proved highly beneficial for feature libraries lacking optimal features.

Conclusions:

• Extended representations of pixel planes (EPP) are effective in improving class separability.
• Image transform-based EPP offer superior performance, especially when optimal features are not yet established.
• EPP provides a valuable approach for enhancing feature representation in machine learning and computer vision tasks.