Numerical accuracy of integral images computation algorithms

In this study, the authors examine the issue of numerical accuracy of different variants of integral image computation algorithms both sequential and parallel (e.g. well-suited for graphics processing units). The experimental study shows that particular arrangements of successive arithmetical operations for parallel integral image computation algorithms as well as a row-column approach can lead up to a substantial numerical accuracy improvement over classic implementations. Such accuracy differences are then experimentally shown, on the example of the image binarisation problem, to be of great significance in terms of quality of final results being obtained by image processing methods utilising integral images as a part of their processing framework. In addition, the authors draw conclusions which can later be used in the formulation of general guidelines for constructing high-quality and time-effective implementations of image processing methods involving integral images.