MA-CRNN: a multi-scale attention CRNN for Chinese text line recognition in natural scenes

Tong, Guofeng; Li, Yong; Gao, Huashuai; Chen, Huairong; Wang, Hao; Xiang, Yang

doi:10.1007/s10032-019-00348-7

Cited by 22 publications

(11 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The authors report that "[l]arger sample heights did not improve recognition accuracy for skew-free text lines"[NK19]. 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].…”

mentioning

confidence: 89%

“…In existing research on OCR methods, one important distinction to be made is between older, character-based methods and more recent, line-based methods: the former seg-ment single characters to recognize them one by one, the latter process whole text lines at once [Ton+20]. While character-based methods are still successfully used in scene text detection models [CSA20], most researchers agree that for recognizing printed text scanned in high resolutions line-based deep neural network (DNN) models are now the preferred state of the art solution [Bre+13;WRP18a;Reu+19b;MWL18;Zho+17].…”

Section: Ocr In the Age Of Deep Learningmentioning

confidence: 99%

“…For example, Bluche et al showed how to extend CTC with attention mechanisms to improve HTR performance [BLM17]. Another attention mechanism that extended BLSTMs was proposed by Tong et al [Ton+20]. One recent advance in the overall quality of CRNN models is the introduction of ensemble voting by Wick et al [Reu+18], which follows an established OCR method of combining multiple classifiers [Han98].…”

Section: Ocr In the Age Of Deep Learningmentioning

confidence: 99%

“…Choosing a line height for a CRNN architecture is a tradeoff between getting more detail and getting better generalization. If a line height is too small, too much information might get lost to differentiate certain characters [Ton+20]. This is especially true for historical typefaces like blackletter where very small details can produce completely different characters.…”

Section: Line Heightmentioning

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: 89%

Section: Ocr In the Age Of Deep Learningmentioning

confidence: 99%

Section: Ocr In the Age Of Deep Learningmentioning

confidence: 99%

Section: Line Heightmentioning

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

“…ey were then developed by other researchers. A CNN is a multilayer neural network which can benefit simultaneously from an automated feature extractor and trainable classifier [45]. In recent years, CNNs have most widely been used in solving machine vision problems such as pattern recognition, object detection, or speech recognition [1].…”

Section: Deep Networkmentioning

confidence: 99%

Recognition of Persian/Arabic Handwritten Words Using a Combination of Convolutional Neural Networks and Autoencoder (AECNN)

Khosravi

Chalechale

2022

Mathematical Problems in Engineering

View full text Add to dashboard Cite

Despite extensive research, recognition of Persian and Arabic manuscripts is still a challenging problem due to the complicated and irregular nature of writing, wide vocabulary, and diversity of handwritings. In Persian and Arabic words, letters are joined together, and signs such as dots are placed above or below letters. In the proposed approach, the words are first decomposed into their constituent subwords to enhance the recognition accuracy. Then the signs of subwords are extracted to develop a dictionary of main subwords and signs. The dictionary is then employed to train a classifier. Since the proposed recognition approach is based on unsigned subwords, the classifier may make a mistake in recognizing some subwords of a word. To overcome this, a new subword fusion algorithm is proposed based on the similarity of the main subwords and signs. Here, convolutional neural networks (CNNs) are utilized to train the classifier. An autoencoder (AE) network is employed to extract appropriate features. Thus, a hybrid network is developed and named AECNN. The known Iranshahr dataset, including nearly 17000 images of handwritten names of 503 cities of Iran, was employed to analyze and test the proposed approach. The resultant recognition accuracy is 91.09%. Therefore, the proposed approach is much more capable than the other methods known in the literature.

show abstract

HCapsNet: A Text Classification Model Based on Hierarchical Capsule Network

2021

Lecture Notes in Computer Science

View full text Add to dashboard Cite

MA-CRNN: a multi-scale attention CRNN for Chinese text line recognition in natural scenes

Cited by 22 publications

References 23 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

Recognition of Persian/Arabic Handwritten Words Using a Combination of Convolutional Neural Networks and Autoencoder (AECNN)

HCapsNet: A Text Classification Model Based on Hierarchical Capsule Network

Contact Info

Product

Resources

About