SwinCoder: A Swin Transformer-based Image Compression Model with Perceptual Optimization For CLIC2025

Daxin Li

Abstract: In this paper, we present a low-complexity image compression model based on the Swin Transformer, which only requires 100kMACs/pixel. We introduce a perceptual optimization strategy by incorporating adversarial training. Experimental results demonstrate that our model improves the perceptual quality of compressed images. This paper is a solution of CLIC2025 challenge and our team name are TestC and RunRun for GPU and CPU tracks, respectively.

Team Name: TestC,RunRun

Submission Number: 3