Data Augmentation Using Bitplane Information Recombination Model.

Area of Science:

• Computer Vision
• Machine Learning
• Data Science

Background:

• Deep learning performance relies heavily on large, high-quality datasets.
• Acquiring sufficient annotated data is a significant challenge in many scientific fields.
• Limited data hinders the training and generalization capabilities of neural networks.

Purpose of the Study:

• To introduce a novel data augmentation technique to address data scarcity in deep learning.
• To propose the bitplane information recombination model (BIRD) for generating diverse training data.
• To demonstrate the effectiveness of BIRD in improving model performance without altering original annotations.

Main Methods:

• The BIRD method decomposes images into bitplanes, capturing hierarchical structural information.
• It employs bitplane extraction, selection, and recombination to create augmented data with varied details.
• This approach is model-agnostic, parameter-free, and integrates seamlessly with existing neural networks.

Main Results:

• BIRD can generate up to 62 times the original training data for 8-bit images.
• Experiments on target detection (DOTA dataset) and image classification (CIFAR-100 dataset) showed significant improvements.
• The method effectively augmented data and enhanced accuracy in both tested applications.

Conclusions:

• The BIRD method offers a powerful solution for data augmentation in deep learning.
• It effectively overcomes data limitations, leading to improved accuracy in computer vision tasks.
• BIRD provides a versatile and efficient approach for enhancing neural network training datasets.