Focus on images: when “less is more” suits Transkribus. Observations on colors and image quality for better text recognition

As an HTR software, Transkribus is primarily designed to extract text from image data and, therefore, working with images is its main focus. This reflection begins with a case study. Given two different reproductions of the same document, one of low quality in black and white, obtained by the digitization of a microfilm, and the other of high quality in color, it is proposed to examine the reaction of Transkribus to two such different images. Although of lower quality, black-and-white images are processed better than color images due to a higher contrast between text and background; moreover, color images introduce additional complexity and noise that may interfere with the accuracy of the HTR. In addition to pointing out the difference in how the software behaves on the two different reproductions, we would like to suggest the opportunity to enhance the images and binarize them to achieve a better transcription.