Dense prediction for text line segmentation in handwritten document images
Quang Nhat Vo, Guee-Sang Lee · 2016
We present a novel approach to segment text lines from handwritten document images. In contrast to existing approaches which mainly use hand-designed features or heuristic rules to estimate the location of text lines, we train a fully convolutional network (FCN) to predict text line structure in document images. By using the FCN, a line map which is a rough estimation of text line is obtained. From this line map, text strings that pass through characters in each text line are constructed. To deal with touching text lines, line adjacency graph (LAG) is used to separate the touching characters into different text strings. The testing result on ICDAR2013 Handwritten Segmentation Contest dataset shows high performance together with the robustness of our system with different types of languages and multi-skewed text lines.