Unraveling Neural Cellular Automata for Lightweight Image Compression

Andreza Maria Coutinho Falcao, IURY ADONES XAVIER DOS SANTOS, Filipe R. Cordeiro

Keywords: image compression, automata, deep learning

Abstract: Neural Cellular Automata (NCA) are computational models inspired by cellular growth, capable of learning complex behaviors through local interactions. While NCAs have been applied to various tasks like image restoration and synthesis, their potential for image compression remains largely unexplored. This paper aims to unravel the capabilities of NCAs for lightweight image compression by introducing a Grid Neural Cellular Automata (GNCA) training strategy. Unlike traditional methods that depend on large deep learning models, NCAs offer a low-cost compact and highly parallelizable alternative with intrinsic robustness to noise. Through experiments on the COCO 2017 dataset, we compare the compression performance of NCAs against JPEG, JPEG-2000 and WebP, using the metrics PSNR, SSIM, and MSE and Compression Rate. Our results demonstrate that NCAs achieve competitive compression rates and image quality reconstruction, highlighting their potential as a lightweight solution for efficient image compression. The code will be available upon acceptance.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12047