Efficient Binarization for Historical Document Analysis

Readability of document images is one core issue when analysing historical documents. One way to improve the readability of those document images is image binarization. By separating the written text from its background, documents degraded by, e.g., stains or faded ink become better readable. Due to the large quantity of available historical document images, this binarization needs to be done efficiently to make this form of processing feasible. In this paper, we present our work in progress on improving the execution performance of a state-of-the-art binarization algorithm by mapping it onto a heterogenous platform. We describe how the algorithm can be divided and computed in parallel on CPU and GPU. The preliminary results, which we report in this paper, suggest that a speedup of 1.4 is possible in comparison to the original algorithm.