A Few-shot Learning Approach for Historical Ciphered Manuscript Recognition

Souibgui, Mohamed Ali; Fornés, Alícia; Kessentini, Yousri; Tudor, Crina

doi:10.1109/icpr48806.2021.9413255

Cited by 12 publications

(16 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…To find the right amount of lines from each set that should be added to the total mix, we perform an experiment of varying the percentages of the BPLL and the DAL, and calculating the SER at each time. The used model for this training is the one presented in [20] which shows a good results for Borg manuscript. As shown in Fig.…”

Section: Htr Resultsmentioning

confidence: 99%

“…Given those "shots", the system must output the transcribed text based on a similarity matrix between the provided alphabet and the text lines. These methods are now used for text [23] and historical ciphered text [20] recognition.…”

Section: Handwritten Ciphered Text Recognitionmentioning

confidence: 99%

“…As can be seen, this cipher is hard to be segmented, mainly because of the frequent symbol overlapping not only between consecutive symbols, but also between the different lines. Following related works [1,20], we have used 273 lines extracted from 16 pages for testing. Note that a pre-processing step (binarization and projections) has been applied to obtain those lines from the full pages manuscripts.…”

Section: Datasetmentioning

confidence: 99%

“…After generating the synthetic Borg lines, a machine learning model should be trained for recognition. For this, we consider two options: A supervised model based on sequence to sequence with attention [14] and a few-shot learning based model [20].…”

Section: Htr Models and Evaluation Metricmentioning

confidence: 99%

“…Few-shot for HTR Since we are using one or few examples of each Borg symbols to generate the data (a few-shot generation), using a few-shot model for the recognition could be suitable. Thus, we choose the approach proposed in [20], a segmentation free (works at line level) method for historical ciphered handwritten text recognition. It consists in inputting a ciphered text line image with an associated alphabet as isolated symbols to a matching model, where one or few examples (usually up to five) of each symbol should be given.…”

Section: Htr Models and Evaluation Metricmentioning

confidence: 99%

See 4 more Smart Citations

One-shot Compositional Data Generation for Low Resource Handwritten Text Recognition

Souibgui¹,

Biten²,

Dey³

et al. 2021

Preprint

Self Cite

View full text Add to dashboard Cite

Low resource Handwritten Text Recognition (HTR) is a hard problem due to the scarce annotated data and the very limited linguistic information (dictionaries and language models). This appears, for example, in the case of historical ciphered manuscripts, which are usually written with invented alphabets to hide the content. Thus, in this paper we address this problem through a data generation technique based on Bayesian Program Learning (BPL). Contrary to traditional generation approaches, which require a huge amount of annotated images, our method is able to generate human-like handwriting using only one sample of each symbol from the desired alphabet. After generating symbols, we create synthetic lines to train state-of-the-art HTR architectures in a segmentation free fashion. Quantitative and qualitative analyses were carried out and confirm the effectiveness of the proposed method, achieving competitive results compared to the usage of real annotated data.

show abstract

Section: Htr Resultsmentioning

confidence: 99%

Section: Handwritten Ciphered Text Recognitionmentioning

confidence: 99%

Section: Datasetmentioning

confidence: 99%

Section: Htr Models and Evaluation Metricmentioning

confidence: 99%

Section: Htr Models and Evaluation Metricmentioning

confidence: 99%

See 3 more Smart Citations