Model-based ruling line detection in noisy handwritten documents

Ruling lines are commonly used to help people write neatly on paper. In document analysis, however, they raise hurdles for the tasks of handwriting recognition or writer identification. In this paper, we model ruling line detection as a multi-line linear regression problem and then derive a globally optimal solution under the Least Squares Error. For performance evaluation, we compute the error statistics on the model attributes and also employ human correction of algorithmic results for performance evaluation, instead of using pixel-level performance measures. We demonstrate the effectiveness of our method on three datasets, including modern and historic document images. Specifically, we obtained 95% accuracy in detecting ruling lines in a modern handwriting dataset with 100 documents. Under an interactive evaluation framework, the new algorithm showed performance gains over one existing approach.

► Model-based pre-printed ruling line detection is proposed.
► New paradigm of performance evaluation is proposed.
► Performance evaluation and comparison is conducted on both modern and historical datasets.
► Human correction is employed to correlate the proposed performance measure.
► Acquire a 95% overall accuracy on a dataset with 100 testing images.