The implicit neural representation (INR) employed in image compression shows high decoding efficiency, yet it requires long encoding times due to the need for the model training tailored to the specific image being coded. Thus, we propose a new image compression scheme leveraging the 2D Gaussian splatting technique to accelerate encoding speed and maintain decoding efficiency. Specifically, we parameterize these Gaussians with key attributes including position, anisotropic covariance, color, and opacity coefficients, totaling 9 parameters per Gaussian. We initialize these Gaussians by sampling points from the image, followed by employing an α- blending mechanism to determine the color values of each pixel. For compact attribute representation, we adopt a K-means based vector quantization approach for anisotropic covariance, color and opacity coefficients. Additionally, we introduce an adaptive dense control methodology to dynamically adjust Gaussian numbers, facilitating automatic point reduction or augmentation. Finally, the position, codebooks and indexes of other attributes are quantized and compressed by the lossless entropy coding. Our experimental evaluation demonstrates that our method achieves faster encoding speeds compared to other INR techniques while exhibiting comparable decoding speeds.
The source code is available via the following link: https://github.com/ppingzhang/2DGS_ImageCompression.
Companion
APSIPA Transactions on Signal and Information Processing Special Issue - Deep Learning-Based Data Compression
See the other articles that are part of this special issue.