This competition aims to motivate research around Super-Resolution (SR) and its application in the specific context of Text Images. Optical Character Recognition (OCR) systems allow one to extract textual information from images. However, when those images lack of resolution, these systems fail. Simple interpolation techniques are very limited at improving the performances. SR approaches aim to enhance the reconstruction process by generating missing details, and are able to provide better shaped images yielding better OCR performances.
A dataset was constructed for this competition in order to evaluate the ability of a given SR system to improve those performances. HR images (High-Resolution) were extracted from French TV video flux, and downsampled by a factor of 2 to create LR images (Low-Resolution). A single frame was extracted for each image (Single Image Super-Resolution). They were manually annotated (transcribed).
The dataset is divided into a training set (training, learning, validation purposes) and a test set :