Radon Cumulative Distribution Transform Subspace Modeling for Image Classification.

Area of Science:

• Computer Vision
• Machine Learning
• Image Processing

Background:

• Image classification is crucial in machine learning but challenged by complex deformations like scaling and translation.
• Existing methods, particularly neural networks, can be computationally intensive and require large labeled datasets.
• Developing robust image classification techniques that handle deformations efficiently is an ongoing research area.

Purpose of the Study:

• To introduce a novel supervised image classification method leveraging the Radon Cumulative Distribution Transform (R-CDT).
• To demonstrate the efficacy of R-CDT in handling various image deformations, simplifying classification tasks.
• To compare the proposed method's performance against state-of-the-art neural networks in terms of accuracy, efficiency, and generalization.

Main Methods:

• Utilizing the Radon Cumulative Distribution Transform (R-CDT) to represent image data in a machine learning-friendly format.
• Implementing a nearest-subspace algorithm within the R-CDT space for classification.
• Evaluating the method's performance on broad classes of image deformation models.

Main Results:

• The R-CDT effectively captures image variations caused by transformations, making classification more tractable.
• The proposed method achieves competitive accuracy compared to state-of-the-art neural networks.
• Significant improvements are observed in computational efficiency (GPU-independent), label efficiency (fewer training samples), and out-of-distribution generalization.

Conclusions:

• The R-CDT-based image classification method offers a computationally efficient and label-efficient alternative to deep learning approaches.
• This method demonstrates robustness to image deformations and strong generalization capabilities.
• The approach is simple to implement, non-iterative, and requires no hyper-parameter tuning, making it broadly applicable.