COIN: COmpression with Implicit Neural representations

TL;DR

COIN encodes images by overfitting a small MLP to map pixel coordinates to RGB values, then transmits the quantized weights as the compressed code. Decoding simply evaluates the MLP at each pixel, enabling progressive decoding and avoiding large decoders. Across Kodak images, COIN outperforms JPEG at low bit-rates without entropy coding and highlights the potential of reframing data compression as model compression. While not state-of-the-art yet, the approach offers a simple, decoder-light alternative with clear paths for improvement via learned weight distributions and meta-learning. This work opens a new direction in neural data compression centered on implicit 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.

Figures (14)

• Figure 1: Compressed implicit neural representations. We overfit an image with a neural network mapping pixel locations $(x, y)$ to RGB values (often referred to as an implicit neural representation). We then quantize the weights $\theta$ of this neural network to a lower bit-width and transmit them.
• Figure 2: Rate distortion plots on the Kodak dataset.
• Figure 3: Model sizes at 0.3bpp.
• Figure 4: Model training on image 15 in the Kodak dataset. Max COIN represents the max PSNR achieved at any point during training.
• Figure 5: Plot of maximum PSNR for networks of the same size (0.3bpp) with different architectures.