We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:


Current browse context:


Change to browse by:

References & Citations


(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Electrical Engineering and Systems Science > Image and Video Processing

Title: COIN: COmpression with Implicit Neural representations

Abstract: We propose a new simple approach for image compression: instead of storing the RGB values for each pixel of an image, we store the weights of a neural network overfitted to the image. Specifically, to encode an image, we fit it with an MLP which maps pixel locations to RGB values. We then quantize and store the weights of this MLP as a code for the image. To decode the image, we simply evaluate the MLP at every pixel location. We found that this simple approach outperforms JPEG at low bit-rates, even without entropy coding or learning a distribution over weights. While our framework is not yet competitive with state of the art compression methods, we show that it has various attractive properties which could make it a viable alternative to other neural data compression approaches.
Comments: Added qualitative comparisons and link to github repo this https URL
Subjects: Image and Video Processing (eess.IV); Computer Vision and Pattern Recognition (cs.CV); Machine Learning (cs.LG)
Cite as: arXiv:2103.03123 [eess.IV]
  (or arXiv:2103.03123v2 [eess.IV] for this version)

Submission history

From: Emilien Dupont [view email]
[v1] Wed, 3 Mar 2021 10:58:39 GMT (1114kb,D)
[v2] Sat, 10 Apr 2021 16:36:00 GMT (9272kb,D)

Link back to: arXiv, form interface, contact.