Abstract: Single Image Super Resolution (SISR) is an ill-posed problem: the lack of an intrinsic notion of “same content” across resolutions motivates its statistical formulation. We begin by clarifying this viewpoint and its relevance to our medical imaging setting. We then present an approach based on the additive refinement of the wavelet coefficients of a classically upsampled image. Building on this formulation, we introduce a super-resolution architecture that implements it by combining two-dimensional discrete wavelet transforms with the Swin Transformer. We conclude by discussing the training strategy and the use of knowledge distillation, along with quantitative and qualitative results on medical datasets.
Il seminario è organizzato dai dottorandi di Matematica e si svolgerà in presenza presso il Dipartimento di Matematica e Fisica, via della Vasca Navale 84 , aula B.
Link identifier #identifier__110082-1Sito web
This post is also available in:
Link identifier #identifier__73190-5
Eng