Dataset Color Quantization: A Training-Oriented Framework for Dataset-Level Compression

Researchers propose Dataset Color Quantization (DCQ), a new framework that compresses visual datasets by reducing color-space redundancy while preserving information crucial for AI model training. The method achieves significant storage reduction across major datasets including CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K while maintaining training performance.

• →DCQ addresses storage challenges of large-scale image datasets in resource-constrained environments by targeting color-space redundancy rather than discarding samples.
• →The framework enforces consistent palette representations across similar images while retaining semantically important colors guided by model perception.
• →Extensive testing across major datasets demonstrates significant storage compression with maintained training performance.
• →The approach offers a scalable solution for dataset-level storage reduction in deep learning applications.
• →DCQ preserves structural details necessary for effective feature learning while achieving aggressive compression ratios.