Window-Based Hierarchical Dynamic Attention for Learned Image Compression

Yuan Li, Wei Gao

Keywords: Dynamic attention, learned image compression, adaptive entropy model.

Abstract: Transformers have been successfully applied to learned image compression (LIC). In fact, dense self-attention is difficult to ignore contextual information that degrades the entropy estimations. To overcome this challenging problem, we incorporate dynamic attention in LIC for the first time. The window-based dynamic attention (WDA) module is proposed to adaptively tune attention based on entropy distribution by sparsifying the attention matrix. Additionally, the WDA module is embedded into encoder and decoder transformation layers to refine attention in multi-scales, hierarchically extracting compact latent representations. Similarly, we propose the dynamic-reference entropy model (DREM) to adaptively select context information. This decreases the difficulty of entropy estimation by leveraging the relevant subset of decoded symbols, achieving an accurate entropy model. To the best of our knowledge, this is the first work employing dynamic attention for LIC and extensive experiments demonstrate the proposed method outperforms the state-of-the-art LIC methods.

Primary Area: other topics in 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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

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: 11674