ICDAR 2015 Competition on Text Image Super-Resolution

Task

Participants are asked to provide a Super-Resolved version of a set of Low Resolution images, for an upscaling factor of 2.

Goal

Provided SR images must yield the best OCR and PSNR scores improvement, compared to basic bicubic interpolated images.

Evaluation Procedure

Text Image Super-Resolution allows to evaluate a given method with 2 different kind of measures: pixel-wise errors and OCR accuracy. Usually, Super-Resolution methods are evaluated using PSNR, MSE or sometimes SSIM measures. They allow to evaluate how well a SR image (result of a SR algorithm from a LR image) is reconstructed compared with an original ideal HR image. Here, we believe that Super-Resolution can increase the letter recognition rate of OCR systems; therefore the evaluation is also based on the capacity of a given method to eventually yield a better OCR score.

As the competition is dedicated to Text Images, a particular attention will be paid to the OCR score enhancement via Super-Resolution for the final evaluation. But to stay in the scope of Super-Resolution, one or more reconstruction quality score like PSNR will be considered as well.

To compute OCR score, Tesseract-OCR 3.02 will be used for sake of reproducibility, as it is a free, open-source, and well-documented software.