Brain-inspired semantic data augmentation for multi-style images

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Data augmentation is crucial for deep learning, but common methods struggle with large style differences between training and testing data.
• Brain-inspired methods offer novel approaches to AI by mimicking biological principles.
• Domain shifts, where data distributions differ, hinder model generalization.

Purpose of the Study:

• To propose a new brain-inspired data augmentation method to enhance deep learning model generalization.
• To address the limitations of existing methods when dealing with significant domain shifts and style variations.
• To improve the robustness and performance of computer vision models on diverse datasets.

Main Methods:

• Improved modeling of Domain Shifts with Uncertainty (DSU).
• Proposed a two-component method: Robust statistics and controlling the Coefficient of variance for DSU (RCDSU) and Feature Data Augmentation (FeatureDA).
• RCDSU uses robust statistics to mitigate outlier influence and controls variance for semantic preservation and increased shift range. FeatureDA augments features with unchanged semantics and increased coverage.

Main Results:

• RCDSU and FeatureDA achieve competitive accuracy on the Photo, Art Painting, Cartoon, and Sketch (PACS) multi-style classification task.
• The combined method demonstrates strong robustness against outliers when Gaussian noise is added to the PACS dataset.
• FeatureDA alone shows excellent results on the CIFAR-100 image classification task.

Conclusions:

• The proposed RCDSU plus FeatureDA method is a novel brain-inspired semantic data augmentation technique.
• This method is suitable for datasets with large style differences between training and testing data, offering implicit robot automation.
• The approach effectively improves model generalization at both style and content levels, enhancing robustness and performance.