A Transformer-Based U-Net Architecture for Fast and Efficient Image Demoireing
Authors
Densen PuthusseryP. S. HrishikeshC. V. Jiji
Abstract
Recently, transformer based deep neural networks have been found useful in solving various image restoration tasks like image denoising, deblurring, deraining etc., producing significant improvement in PSNR and SSIM over CNN based techniques on benchmark datasets. These networks have effectively addressed quadratic computational complexity issue with increasing image resolution by making use of novel self attention strategies on local image windows. In this paper, we propose a fast and efficient UNet based architecture using transformer modules for the image demoireing task. The proposed architecture is computationally very efficient as the transformer blocks perform non-overlapping window-based self-attention instead of global self attention.
Why we worked on this
This is the most-cited paper of the five, and the reason is efficiency, not novelty. Transformers improved image restoration and made it too expensive to run; the contribution here is an architecture that keeps the quality while staying cheap enough to deploy. That trade — accuracy against what a customer can actually afford to run — is the one we make on nearly every engagement.
Where to start
The AI Modernization Assessment is two weeks and $5,000. Not a code audit. You get a prioritized roadmap, effort and cost ranges, and a written list of the workflows where AI would add risk without adding value.
Book the assessment