Lossy compression with implicit neural representations utilizing chroma subsampling

Abstract

Lossy compression is the process in which the file size of information is reduced while retaining the overall fidelity of the information. It is utilized to reduce the amount of data for storage and transmission. For specifically lossy image compression, lots of data can be removed while keeping the image nearly identical due to the fact that the human visual system is inexact. Lossy image compression algorithms such as the Joint Photographic Experts Group (JPEG) are utilized to accomplish this goal. Recently, a new form of signal representation has been proposed known as the Implicit Neural Representation (INR). Instead of explicitly defining the compression process, a multi­layer perceptron (MLP) network learns to memorize a signal storing the information as parameters of a MLP as opposed to the image itself. While INRs have shown promise, they have yet to become viable compression methods because traditional methods are faster and more accurate. Many methods have been utilized to further improve the compression ratio of Implicit Neural Representations such as quantization and entropy coding. In this thesis, we explore an image processing technique known as Chroma Sub­sampling to improve the compression quality of INRs. This method takes advantage of the fact that the human visual system is more susceptible to details in brightness rather than details in colors. Additionally, we will also introduce various architectures and techniques to train a chroma subsampled image into an INR more effectively.