One of the most important steps in a handwriting recognition system is text-line and word segmentation. But, this step is made difficult by the differences in handwriting styles, problems of skewness, overlapping and touching of text and the fluctuations of text-lines. It is even more difficult for ancient and calligraphic writings, as in Arabic manuscripts, due to the cursive connection in Arabic text, the erroneous position of diacritic marks, the presence of ascending and descending letters, etc. In this work, we propose an effective segmentation of Arabic handwritten text into text-lines and words, using deep learning. For text-line segmentation, we used an RU-net which allows a pixel-wise classification to separate text-lines pixels from the background ones. For word segmentation, we resorted to the text-line transcription, as we have not got a ground truth at word level. A BLSTM-CTC (Bidirectional Long Short Term Memory followed by a Connectionist Temporal Classification) is then used to perform the mapping between the transcription and text-line image, avoiding the need of the input segmentation. A CNN (Convolutional Neural Network) precedes the BLST-CTC to extract the features and to feed the BLSTM with the essential of the text-line image. Tested on the standard KHATT Arabic database, the experimental results confirm a segmentation success rate of no less than 96.7% for text-lines and 80.1% for words.
Page stream segmentation into single documents is a very common task which is practiced in companies and administrations when processing their incoming mail. It is not a straightforward task because the limits of the documents are not always obvious, and it is not always easy to find common features between the pages of the same document. In this paper, we seek to compare existing segmentation models and propose a new segmentation one based on GRUs (Gated Recurrent Unit) and an attention mechanism, named AGRU. This model uses the text content of the previous page and the current page to determine if both pages belong to the same document. So, due to its attention mechanism, this model is capable to recognize words that define the first page of a document. Training and evaluation are carried out on two datasets: Tobacco-800 and READ-Corpus. The former is a public dataset on which our model reaches an F1 score equal to 90%, and the later is private for which our model reaches an F1 score equal to 96%.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.