AI Paper Atlas

Research map / Continual learning / Online and replay-based

Expandable, Compressible, Mineable: Open-World Thermal Infrared Image Restoration

Pu Li, Huafeng Li, Yafei Zhang, Wen Wang, Neng Dong, Jie Wen

ICML 2026 · Continual learning › Online and replay-based · approach: Dynamic expansion and isolation

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use group-isolated expansion, SEP and SKMM together instead of a single shared backbone (ablation)ECMRNet · HM-TIR averagePSNR: 26.21 -> 30.36Improved
add Structural Entropy Pruning (ablation)ECMRNet · HM-TIR averagePSNR / Params: Params 27.51M -> 19.40M, PSNR 30.51 -> 30.36Improved
add SKMM knowledge mining for compound degradations (ablation)ECMRNet · HM-TIR compound degradationsPSNR: 25.44 -> 26.36Improved
add SKMM (ablation)ECMRNet · HM-TIR compound degradationsFLOPs: 76.12G -> 297.97GHurt
raise SEP frequency threshold rho from 0.10 to 0.20 (ablation)SEP · HM-TIR contrastPSNR: 37.46 -> 36.34Hurt
lower SEP frequency threshold rho from 0.10 to 0.05 (ablation)SEP · HM-TIR contrastParams / PSNR: PSNR 37.46 -> 37.54, Params 2.95M -> 3.59MMixed
increase SEP sample pool size N from 100 to 400 (ablation)SEP · HM-TIR contrastPSNR: 37.46 -> 37.47No clear effect
use SEP instead of magnitude pruning or sparsity-regularized pruning (ablation)channel-group pruning · HM-TIR (no fine-tuning, avg)PSNR: MP 18.90, SRP 15.75 -> SEP 25.71Improved
apply SEP to longer task sequence (dead pixels task) (ablation)ECMRNet · HM-TIR dead pixelsPSNR / Params: PSNR 45.34 -> 45.10, Params 4.07M -> 2.37MImproved
use ECMRNet instead of all-in-one restorers and continual-learning baselinesTIR restoration · HM-TIR averagePSNR: 28.77 -> 30.36 (vs best baseline MoCE-IR)Improved
use regularization-based continual learning (LwF, EWC, MAS, SI) on NAFNetNAFNet · HM-TIR averagePSNR: NAFNet 26.39 -> LwF 22.12, EWC 21.48, MAS 22.34, SI 22.03Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy