Positive Data Augmentation Based on Manifold Heuristic Optimization for Image Classification.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Data augmentation is vital for insufficient training data, particularly for positive samples.
• Existing methods often neglect feature distribution optimization and rely on neural network feedback.

Purpose of the Study:

• To develop a practical, distribution-preserving data augmentation pipeline.
• To augment positive samples while maintaining alignment with original data distribution.

Main Methods:

• Proposed the Manifold Heuristic Optimization Algorithm (MHOA) inspired by the manifold hypothesis.
• Augmented samples by exploring low-dimensional Euclidean space around object contour pixels.
• Optimized feature indicator fidelity to the original data manifold and retained samples with aligned feature statistics.

Main Results:

• Significantly improved image classification accuracy across various neural networks.
• Outperformed state-of-the-art data augmentation methods, especially for Gaussian-distributed feature indicators.
• Demonstrated superior performance driven by a focused search space on key feature pixel neighborhoods.

Conclusions:

• The MHOA pipeline offers an effective strategy for distribution-preserving data augmentation.
• This approach enhances model performance by optimizing feature distribution fidelity.
• The method shows particular promise for datasets with Gaussian feature distributions.