Efficient, Lexicon-Free OCR using Deep Learning

Namysł, Marcin; Konya, Iuliu

doi:10.48550/arxiv.1906.01969

Cited by 2 publications

(7 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This 32 px limit is supported by Tong et al, who report that a height of 32 px produced the same accuracy as 48 px and 64 px, but that 16 px was too little to correctly recognize small characters[Ton+20]. On the other hand, Namsyl and Konya also state that "relatively long, free-form text lines [warrant] the use of taller samples"[NK19]. This leads us to the situation with current state of the art (non-HTR) CRNN systems geared towards historical typefaces, where line heights of 40 px [MLK20; MGK19] and, more commonly, 48 px [WRP18a; Bre+13; KHK19] are to be found 4 .…”

mentioning

confidence: 82%

“…We omitted P = 0 because preliminary results showed that this did not work well at all. Following (G3) and the fact that at least 10 layers have been used in fully convolutional networks for text recognition [NK19], we focus more on depth (K) than on width (N and R). This is now a general practice in deep learning, as the "depth of a neural network is exponentially more valuable than the width of a neural network" [WR17].…”

Section: Findings ψmentioning

confidence: 99%

“…For example, [SSP03] note that "[a]ffine distortions greatly improved our results," but "our best results were obtained when we used elastic deformations." Similarly, [NK19] report "that non-linear distortions can [...] reduce the error rate of models". Accordingly, we take a closer look at two forms of non-linear distortions that we apply on the whole line image (not single characters):…”

Section: Geometric Deformationsmentioning

confidence: 99%

“…For example, Michael et al give a detailed description on their data augmentation as consisting of "affine transformations, dilation and erosion, as well as elastic and grid-like distortions"[MWL18]. Similarly, Namysl and Konya use "Gaussian smoothing, perspective distortions, morphological filtering, downscaling, additive noise, and elastic distortions" as well as a new augmentation called "alpha compositing"[NK19]. Martinek et al create training data synthetically by concatenating letter symbols [MLK20; MLK19], but it is not clear how these methods fare in comparison to or in combination with other data augmentation methods.…”

mentioning

confidence: 99%

See 3 more Smart Citations

On the Accuracy of CRNNs for Line-Based OCR: A Multi-Parameter Evaluation

Liebl,

Burghardt

2020

Preprint

View full text Add to dashboard Cite

We investigate how to train a high quality optical character recognition (OCR) model for difficult historical typefaces on degraded paper. Through extensive grid searches, we obtain a neural network architecture and a set of optimal data augmentation settings. We discuss the influence of factors such as binarization, input line height, network width, network depth, and other network training parameters such as dropout. Implementing these findings into a practical model, we are able to obtain a 0.44% character error rate (CER) model from only 10,000 lines of training data, outperforming currently available pretrained models that were trained on more than 20 times the amount of data. We show ablations for all components of our training pipeline, which relies on the open source framework Calamari.

show abstract

mentioning

confidence: 82%

Section: Findings ψmentioning

confidence: 99%

Section: Geometric Deformationsmentioning

confidence: 99%

mentioning

confidence: 99%

See 2 more Smart Citations

On the Accuracy of CRNNs for Line-Based OCR: A Multi-Parameter Evaluation

Liebl,

Burghardt

2020

Preprint

View full text Add to dashboard Cite

show abstract

“…The problem of Optical Character Recognition (OCR) has been the scope of research for many years [1]- [3] due to the need for an efficient method to digitize printed documents, prevent their loss and gradual unavoidable wear, as well as increase their accessibility and portability. The challenges that face Arabic OCR systems stem from the cursive and continuous nature of Arabic scripts.…”

Section: Introductionmentioning

confidence: 99%

An Efficient Language-Independent Multi-Font OCR for Arabic Script

Osman¹,

Zaghw²,

Hazem³

et al. 2020

Preprint

View full text Add to dashboard Cite

Optical Character Recognition (OCR) is the process of extracting digitized text from images of scanned documents. While OCR systems have already matured in many languages, they still have shortcomings in cursive languages with overlapping letters such as the Arabic language. This paper proposes a complete Arabic OCR system that takes a scanned image of Arabic Naskh script as an input and generates a corresponding digital document. Our Arabic OCR system consists of the following modules: Pre-processing, Word-level Feature Extraction, Character Segmentation, Character Recognition, and Post-processing. This paper also proposes an improved font-independent character segmentation algorithm that outperforms the state-of-the-art segmentation algorithms. Lastly, the paper proposes a neural network model for the character recognition task. The system has experimented on several open Arabic corpora datasets with an average character segmentation accuracy 98.06%, character recognition accuracy 99.89%, and overall system accuracy 97.94% achieving outstanding results compared to the state-ofthe-art Arabic OCR systems.

show abstract

Efficient, Lexicon-Free OCR using Deep Learning

Cited by 2 publications

References 0 publications

On the Accuracy of CRNNs for Line-Based OCR: A Multi-Parameter Evaluation

On the Accuracy of CRNNs for Line-Based OCR: A Multi-Parameter Evaluation

An Efficient Language-Independent Multi-Font OCR for Arabic Script

Contact Info

Product

Resources

About