Reproducible Research in Document Analysis and Recognition

With reproducible research becoming a de facto standard in computational sciences, many approaches have been explored to enable researchers in other disciplines to adopt this standard. In this paper, we explore the importance of reproducible research in the field of document analysis and recognition and in the Computer Science field as a whole. First, we report on the difficulties that one can face in trying to reproduce research in the current publication standards. These difficulties for a large percentage of research may include missing raw or original data, a lack of tidied up version of the data, no source code available, or lacking the software to run the experiment. Furthermore, even when we have all these tools available, we found it was not a trivial task to replicate the research due to lack of documentation and deprecated dependencies. In this paper, we offer a solution to these reproducible research issues by utilizing container technologies such as Docker. As an example, we revisit the installation and execution of OCRSpell which we reported on and implemented in 1994. While the code for OCRSpell is freely available on github, we continuously get emails from individuals who have difficulties compiling and using it in modern hardware platforms. We walk through the development of an OCRSpell Docker container for creating an image, uploading such an image, and enabling others to easily run this program by simply downloading the image and running the container.