Invariance encoding in sliced-Wasserstein space for image classification with limited training data.

Area of Science:

• Computer Science
• Machine Learning
• Image Processing

Background:

• Deep convolutional neural networks (CNNs) excel in image classification but struggle with limited training data.
• Traditional data augmentation for CNNs can be computationally expensive and not always effective.

Purpose of the Study:

• To propose a novel mathematical augmentation strategy for image classification.
• To address the limitations of data augmentation in deep learning models, especially with scarce data.
• To offer a computationally efficient and parameter-free alternative.

Main Methods:

• Developed a nearest subspace classification model augmented in sliced-Wasserstein space.
• Utilized the mathematical properties of the Radon Cumulative Distribution Transform (R-CDT) for augmentation.
• Implemented a non-iterative, parameter-free approach.

Main Results:

• Demonstrated superior classification accuracy compared to deep CNNs with data augmentation on limited datasets.
• Achieved significant advantages in computational complexity and efficiency.
• The proposed method proved particularly effective in low-data regimes.

Conclusions:

• The mathematical augmentation using R-CDT offers a simple, effective, and computationally efficient solution for image classification.
• This method provides a viable alternative to traditional data augmentation, especially when training data is limited.
• The approach requires no parameter tuning and is readily available via Python code.