The dataset consists of images corrupted by motion blur together with corresponding high-quality images from two different samples, one of thin sectioned kidney tissue and one of a calibration grid. The data was collected using a MiniTEM microscope (Vironova AB). The motion corrupted images are created by moving the sample under the microscope. Each low-quality (motion blurry) imaging sequence has corresponding high-quality images (captured by stopping the microscope at each position in the sequence). The high-quality frames have a size of 2048 x 2048 pixels with an overlap of 50% between adjacent frames. The low-quality (motion blurry) frames are captured with a size of 1024x1024 with the same motion direction (approximately vertically upwards). All images were captured at a field of view of 32μm, and with a per image exposure time of 15ms and stored as 16 bit tiff files. Both samples are imaged with the same settings and have four imaging sequences each. The dataset contains the raw image files as well as a partitioning into training, validation and testing. For these images, five low-quality images have been registered to each high-quality image. For the five registered images the intersection of all is cropped and stored. 1 of the 4 imaging sequences are chosen as the test set and the last part of another of the imaging sequences as a validation set. The rest is put in the training set. Folder Structures: Raw data: Raw data is the unprocessed data and each sample folder contains 4 image sequences. In each of these folders low-quality (motion blurry) images are stored in folder “Low” and corresponding high-quality images are stored in “GT” TrainValTest: TrainValTest consist of data where the low-quality frames have been registered to the high-quality frames and divided into a training, validation and test set. Each of the Train, Val, Test folders contains 3 subfolders. “Low” contains folders names the same as the files in “GT” where each folder contains five low-quality (motion blurry) images, registered the that corresponding high-quality image. “GT” contains the corresponding high-quality images down sampled to the same spatial size as the low-quality images. “GT_hr” contains the same images as “GT” but not down sampled.

Financially supported by the Swedish Foundation for Strategic Research (grant BD150008), the European Research Council (grant ERC2015CoG 682810), and the Uppsala University AI4Research initiative. Data collected at, and in collaboration with, Vironova AB, Stockholm, Sweden