AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Hybrid Spectral Denoising Transformer with Guided Attention

Zeqiang Lai, Chenggang Yan, Ying Fu

ICCV 2023 · Architectures and efficiency › Attention and transformers · approach: Efficient image transformers · 41 citations (OpenAlex)

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
add spectral self-attention across bands to a conv baseline (ablation)HSDT-M · ICVL Gaussian sigma=50PSNR: 38.28 -> 41.34Improved
add learnable-query guidance to spectral attention (ablation)HSDT-M · ICVL Gaussian sigma=50PSNR: 41.34 -> 41.46Improved
add self-modulated FFN (ablation)HSDT-M · ICVL Gaussian sigma=50PSNR: 41.46 -> 41.62Improved
replace Conv3D with S3Conv (ablation)HSDT-M · ICVL Gaussian sigma=50PSNR: 41.62 -> 41.82Improved
use GSSA instead of MDTA, QRU, GSA/MS-MSA or MGSA (ablation)HSDT-M · ICVL Gaussian sigma=50PSNR: 41.03 -> 41.82 vs MDTA; 39.74 -> 41.82 vs MGSAImproved
use S3Conv instead of previous separable conv Sep3D (ablation)HSDT-M · ICVL Gaussian sigma=50PSNR: 41.44 -> 41.82Improved
implement GSSA with depthwise conv instead of matmul (ablation)GSSA · ICVL Gaussian sigma=50runtime (s): 0.60 -> 0.47 s, PSNR 41.82 unchangedImproved
sequential rather than parallel spatial/spectral convs in S3Conv (ablation)S3Conv · ICVL Gaussian sigma=50PSNR: 41.82 -> 41.58Hurt
HSDT-L versus SST on impulse and mixture noise structure metricsHSDT-L · ICVL complex noise (mixture)SAM: SAM 0.101 vs SST 0.067Hurt

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