Distribution-preserving data augmentation.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning success relies on large datasets, but creating labeled data is challenging and expensive.
• Data augmentation methods increase dataset size and variety by creating image variations.
• Existing color-based augmentation methods often fail to produce plausible image variations.

Purpose of the Study:

• To propose a novel distribution-preserving data augmentation method for creating plausible image variations.
• To improve the generalization and performance of deep learning models through enhanced data augmentation.

Main Methods:

• Developed a novel data augmentation technique that preserves the original image's color distribution.
• Defined a regularized density decreasing direction to shift pixel colors towards distribution tails.
• Applied the method to image classification and segmentation tasks using transfer learning.

Main Results:

• The proposed distribution-preserving method significantly outperforms existing data augmentation techniques.
• Demonstrated superior performance on benchmark datasets including UC Merced Land-use, Intel Image Classification, and Oxford-IIIT Pet.
• Achieved improved accuracy in both classification and segmentation tasks.

Conclusions:

• The novel distribution-preserving data augmentation method effectively generates plausible image variations.
• This approach offers a superior alternative to traditional color-based augmentation methods.
• The method shows strong potential for improving deep learning model performance in various computer vision applications.