Overfitted image coding at reduced complexity

Théophile Blard,Théo Ladune,Pierrick Philippe,Gordon Clare,Xiaoran Jiang,Olivier Déforges

2024-03-18

Abstract:Overfitted image codecs offer compelling compression performance and low decoder complexity, through the overfitting of a lightweight decoder for each image. Such codecs include Cool-chic, which presents image coding performance on par with VVC while requiring around 2000 multiplications per decoded pixel. This paper proposes to decrease Cool-chic encoding and decoding complexity. The encoding complexity is reduced by shortening Cool-chic training, up to the point where no overfitting is performed at all. It is also shown that a tiny neural decoder with 300 multiplications per pixel still outperforms HEVC. A near real-time CPU implementation of this decoder is made available at

Image and Video Processing

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to reduce the encoding and decoding complexity of over - fitting image codecs (such as Cool - chic), while maintaining or improving the compression performance. Specifically, the paper explores the following points:

1. **Reducing decoding complexity**: By simplifying the architecture of the Cool - chic decoder, the decoding performance at different complexity levels is studied. The paper shows that even at an extremely low complexity of 300 multiplications per pixel, Cool - chic can still outperform the traditional HEVC codec.

2. **Increasing decoding speed**: A fast CPU implementation is provided, demonstrating that efficient decoding speed can be achieved even without dedicated hardware. Experimental results show that even for the most complex decoder, its ARM part can complete decoding in only 50 milliseconds.

3. **Reducing encoding complexity**: The impact of shortening the training time on the encoding complexity is studied, and a non - over - fitting (N - O) version of Cool - chic is proposed to address the problem of extremely high encoding complexity. The N - O Cool - chic completes encoding through a single forward pass, greatly reducing the encoding time but sacrificing some compression performance.

### Main contributions

- **Lightweight decoder**: The paper shows how to reduce the decoding complexity from 2,300 multiplications per pixel to 300 multiplications per pixel by simplifying the decoder architecture, while maintaining or even exceeding the compression performance of HEVC.
- **Fast decoding implementation**: A fast CPU implementation is provided, demonstrating that the decoding speed can still meet the requirements of practical applications even at low complexity.
- **Reducing encoding complexity**: By shortening the training time and introducing the non - over - fitting version of Cool - chic, the encoding complexity is significantly reduced, making it more suitable for real - time encoding scenarios.

### Experimental results

- **Rate - distortion performance**: Experiments on the Kodak dataset show that Cool - chic can still outperform HEVC in the lightest - weight setting (300 multiplications per pixel), and is comparable to VVC in the most complex setting (2,300 multiplications per pixel).
- **Decoding speed**: On an AMD EPYC 7282 CPU, even for the most complex decoder, its ARM part only takes 50 milliseconds.
- **Encoding complexity**: By shortening the training time, the encoding complexity can be reduced from 10^9 multiplications per pixel to 10^7 multiplications per pixel, with a performance degradation of less than 10%. The N - O Cool - chic further reduces the encoding complexity to 160 kMAC per pixel, approaching the level of other auto - encoder codecs.

### Conclusion

The paper successfully shows how to significantly reduce the encoding and decoding complexity of Cool - chic while maintaining high performance. These improvements make Cool - chic more suitable for resource - constrained devices and real - time application scenarios.