Taylor series based dual-branch transformation for learned image compression

Abstract

Nonlinear transformations (NTs) are a crucial element of learned image compression (LIC) as they eliminate correlations from images. However, there has been limited focus on developing efficient NTs. This paper aims to propose a design approach for efficient NTs. In this paper, firstly, we model the current NT methods with a unified formulation and find that the difference in the NTs lies in the representation of the scaling function. Subsequently, we analyze the essential elements of designing scaling functions from the perspective of the Taylor series of derivable functions to satisfy the conditions of the Taylor expansion. Lastly, we develop an efficient two-branch nonlinear transformation architecture based on the Taylor series of sinusoidal functions. We compare the performance of various existing nonlinear transformations in the same backbone network. The experimental results demonstrate that our method achieves advanced performance among existing NT methods without a significant increase in encoding and decoding time. For instance, our method achieves a 20.24% BD-rate reduction over BPG444 on the Kodak dataset. The generic capability of our method is validated through experimental results on multiple backbone architectures.