Related Experiment Video
Updated: Jul 7, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
ATD: Improved Transformer With Adaptive Token Dictionary for Image Restoration.
Summary
We introduce Adaptive Token Dictionary (ATD), a novel transformer architecture for image restoration. ATD enables efficient global dependency modeling with linear complexity, achieving state-of-the-art results in super-resolution and denoising.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Deep Learning
Background:
- Transformers excel in image restoration but face computational challenges.
- Quadratic complexity of self-attention limits receptive fields in existing models.
Purpose of the Study:
- To develop a transformer-based image restoration model with linear complexity.
- To enable effective global dependency modeling for improved performance.
Main Methods:
- Proposed Adaptive Token Dictionary (ATD) with a learnable token dictionary.
- Introduced token dictionary cross-attention (TDCA) for feature enhancement.
- Utilized category information for feature grouping and feed-forward network integration.
Main Results:
- ATD and ATD-light achieve state-of-the-art performance on super-resolution benchmarks.
- ATD-U demonstrates effectiveness in image denoising and JPEG artifact removal.
- Models show significant quantitative and qualitative improvements.
Conclusions:
- ATD offers an efficient solution for transformer-based image restoration.
- The proposed methods effectively model global dependencies and enhance feature fusion.
- ATD variants are versatile for various image restoration tasks.
