ScoreReader: A handwritten score recognition toolkit for examination papers

We developed a ScoreReader, a toolkit that a handwritten score recognition for examination papers based on the region of interest (ROI) filtering. ScoreReader can support automatic calculation of the total score for each examination paper and reduce teachers’ workload to help them correct papers efficiently. The toolkit exploits the differences in HSV (Hue, Saturation, and Value) between handwritten scores and background to extract candidate regions accurately. In order to avoid heuristic image segmentation, ScoreReader provides a segmentation-free YOLO recognition approach. In this paper, we evaluate the toolkit on an extended mixed dataset of isolated, double-digit, and connected digits using NIST as the base dataset, and the experimental results demonstrate that the toolkit achieves an accuracy of 95.2% for the computation of handwritten scores on the volume, which is satisfactory performance overall.