Universal Metadata Repository for Document Analysis and Recognition

Document Analysis and Recognition (DAR) has two main objectives, first the analysis of the physical structure of the input image of the document, which should lead to the correct identification of the corresponding different homogeneous components and their boundaries in terms of XY coordinates. Second, each of these homogeneous components should be recognized in such a way that, if it is a text image, consequently this image should be recognized and translated into an intelligible text. DAR remains one of the most challenging topics in pattern recognition. Indeed, despite the diversity of the proposed approaches, techniques and methods, results remain very weak and away from expectations especially for several categories of documents such as complex, low quality, handwritten and historical documents. The complex structure and/or morphology of such documents are behind the weakness of results of these proposed approaches, techniques and methods. One of the challenging problems related to this topic is the creation of standard datasets that can be used by all stakeholders of this topic such as system developers, expert evaluators, and users. In addition, another challenging problem is how one could take advantages of all existing datasets that unfortunately are dispersed around the world without knowing, most of the times, any information about their locations and the way to reach them. As an attempt to solve the two mentioned above problems, we propose in this paper a Universal Datasets Repository for Document Analysis and Recognition (UMDAR) that has, in fact, a twofold advantage. First, it can help dataset creators to standardize their datasets and making them accessible to the research community once published on the proposed repository. Second, it can be used as a central which bridges in a smart manner between datasets and all DAR stakeholders.