Perceptually optimised Cool-chic for CLIC 2025

Pierrick Philippe, Théo Ladune, Gordon Clare, Félix Ernest Henry

Abstract:

Neural image compression has proven to be highly
effective compared to conventional approaches such as JPEG,
HEVC or the latest standard VVC.

Recently low complexity neural image coding using overfitting
has proven to reach the level of performance of VVC. Indeed, at
the CLIC 2024 image challenge, Cool-chic demonstrated to be
perceptually equivalent to the VVC performance at the 3 target
bitrates. The decoding complexity was limited, with less than
2200 operations per pixel, permitting decoding on any legacy
CPU.

For this CLIC 2025 candidate, it is proposed to improve Cool-
chic quality using a perceptually driven distortion metric and the
addition of a random noise.

This paper describes the approach and presents a preliminary
subjective evaluation that demonstrates the effectiveness of the
solution: 50% rate reduction is demonstrated with the proposed
distortion metric. The decoder complexity is reduced to approxi-
mately 1700 operations per pixel, it is written in C language and
it is operated on CPU.

All contributions are made open-source

Team Name: CoolChic

Submission Number: 9