CutSharp: A Simple Data Augmentation Method for Learned Image Compression
Primary Area: general machine learning (i.e., none of the above)
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: learned image compression, data augmentation
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We motivate the effective utilization of data and propose an augmentation technique, called CutSharp, for learned image compression tasks.
Abstract: Learned image compression (LIC) methods have demonstrated superior rate$-$distortion performance, compared to traditional ones. Previous studies on LIC have mainly focused on models, consisting of analysis/synthesis transformations and entropy models. Unfortunately, the importance of $data$ has usually been neglected when training LIC models. In this paper, we introduce block-wise RGB standard deviation as a measure for estimating the compression-related difficulty of images. Next, we emphasize the significance of effective data utilization for LIC by demonstrating that models trained on a certain subset of data, constructed according to the block-wise RGB standard deviation, can achieve superior rate$-$distortion performance to models trained on the entire data. Inspired by this observation, we propose a simple data augmentation technique for LIC, coined CutSharp, which enhances image sharpness within an arbitrary region. Our proposed augmentation consistently improves rate$-$distortion performance on the Kodak and CLIC validation dataset. We hope that our work will encourage further research in data-centric approaches for LIC.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 2288
Loading