Reproduce “Channel-wise Feature Decorrelation for Enhanced Learned Image Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Farhad Pakdaman, Moncef Gabbouj
Researchers from Tampere University developed a channel-wise feature decorrelation loss to improve learned image compression, achieving up to an 8.06% BD-Rate reduction (PSNR) on Cheng's method while adding no computational overhead during inference. This training-time optimization guides deep neural networks to learn more diverse and less correlated latent features, enhancing compression efficiency.
Key insights
- Standard RDO training in Learned Compression leads to correlated latent features, indicating that the deep neural network's capacity is not fully utilized.
- Guiding the latent features to learn a more diverse and less redundant set of representations can significantly enhance compression efficiency.
- Optimizing feature diversity through regularization at training time can improve performance without adding any computational overhead during inference.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.