Image Super-Resolution · Microscopy Images

Image Super-Resolution for Microscopy Images

We gave a small workshop on image super-resolution by applying it to microscopy images.

Dataset Microscopy Images from Zeiss which were downsampled 2x/4x to create the low-resolution versions.

Framework PyTorch · MMagic

Hardware A100 GPU

Task Image Super-Resolution · Microscopy Images · Swin Transformer

Problem Statement

We gave a small workshop as a part of consulting work with Zeiss on image super-resolution by applying it to microscopy images. We also demonstrated how we select frameworks and our mlops practices for training, evaluating the models and sharing the trained models.

Method Overview

We use the SwinIR architecture for image super-resolution, which is based on the Swin Transformer. The model is trained to learn the mapping from low-resolution to high-resolution images using a combination of pixel-wise loss and perceptual loss.

Visual Results

Original image on the left, downsampled image in the middle, and upsampled image on the right.

Quantitative Results

Model	Upsampler	PSNR	SSIM
SwinIR	PixelShuffle	43.20	88.13
SwinIR	PixelShuffle-Direct	42.45	87.90
DinoV2-L	Nearest+Convolution	42.10	87.73

Workshop for image super-resolution on microscopy images.
Workshop on MLOps with mlflow and hyperparameter tuning using grid-search.