Self-Supervised Image Denoising Using Transformers
Name
Pavel Chizhov
Abstract
Self-supervised image denoising is a computer vision task that implies image noise removal without access to clean data. This problem is critical to many domains, such as medical imaging, where clean images are often unobtainable. The absence of the true signal determines the main challenge, therefore self-supervised image denoising demands for specific model engineering. Modern solutions for this task mainly rely on convolutional neural networks, and there has been very limited research on the applications of rapidly developing transformer models to this task. To close this research gap, we adopt a ready-made transformer-based image restoration model for self-supervised image denoising and compare it to the convolutional counterparts. Apart from that, we propose a novel transformer autoencoder architecture, which not only shows more stable performance regardless of the noise type but also is the first model to prove the concept of zero-convolution end-to-end network for self-supervised image denoising. This work highlights the advantages and limitations of transformers in self-supervised image denoising and provides a conceptual foundation for further development in the field.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Mikhail Papkov
Defence year
2023