Ultra Low Bitrate Learned Image Compression by Selective Detail Decoding
Abstract
Neural network-based learned image compression has a special feature in that a differentiable image quality index can be used as a loss function directly, and a decoder and an encoder can be optimized by the quality index through end-to-end learning. From a perceptual view, we hypothesized that there were detailed important parts in pictures. For those parts, we applied an additional decoder and weighted loss function to achieve both low bitrate image compression and perceptual quality. Furthermore, our approach can automatically determine which region an additional decoder will take for an input image. Experiments visually showed that the proposed method can recognize important parts, such as text and faces, and we show that our method can decode images more clearly than the simple MS-SSIM training model.
Cite
Text
Akutsu et al. "Ultra Low Bitrate Learned Image Compression by Selective Detail Decoding." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020. doi:10.1109/CVPRW50498.2020.00067Markdown
[Akutsu et al. "Ultra Low Bitrate Learned Image Compression by Selective Detail Decoding." IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2020.](https://mlanthology.org/cvprw/2020/akutsu2020cvprw-ultra/) doi:10.1109/CVPRW50498.2020.00067BibTeX
@inproceedings{akutsu2020cvprw-ultra,
title = {{Ultra Low Bitrate Learned Image Compression by Selective Detail Decoding}},
author = {Akutsu, Hiroaki and Suzuki, Akifumi and Zhong, Zhisheng and Aizawa, Kiyoharu},
booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops},
year = {2020},
pages = {524-528},
doi = {10.1109/CVPRW50498.2020.00067},
url = {https://mlanthology.org/cvprw/2020/akutsu2020cvprw-ultra/}
}