Adaptive Script-Independent Text Line Extraction

Abstract

SUMMARYIn this paper, an adaptive block-based text line extraction algorithm is proposed. Three global and two local parameters are defined to adapt the method to various handwritings in different languages. A document image is segmented into several overlapping blocks. The skew of each block is estimated. Text block is de-skewed by using the estimated skew angle. Text regions are detected in the de-skewed text block. A number of data points are extracted from the detected text regions in each block. These data points are used to estimate the paths of text lines. By thinning the background of the image including text line paths, text line boundaries or separators are estimated. Furthermore, an algorithm is proposed to assign to the extracted text lines the connected components which have intersections with the estimated separators. Extensive experiments on different standard datasets in various languages demonstrate that the proposed algorithm outperforms previous methods.

Text line segmentation using a fully convolutional network in handwritten document images

“…Also, in this research, we follow the standard of the ICDAR2013 Handwritten Segmentation Contest data set, which only considers the parallel text lines. Another possible extension is the extraction of text string in the form of crossed text lines or complex aligning of text lines [4]. This could be done by creating a new set of line map ground truth and adapting the text string extraction step to the case of crossed text lines that contains several line segments with different angles, which can be considered as another future extension of this work.…”

“…With one line skeleton L i , we only consider the line skeletons L k on the right side that have the similar direction with L i . The merging is continued until there are no pair of line skeletons satisfying the condition in (4). LAG [11] is first introduced as an efficient way to represent CCs in printed music score images.…”

“…The algorithms in this category vary from the way to represent and group the local units. Hough transform-based method [3,4] treats the text units as partitioned blocks of CCs. At first, these CCs are classified to small, large, and normal sized CCs by the average height and width of all CCs.…”