ABSTRACT
End-to-end Optical Character Recognition (OCR) systems are heavily used to convert document images into machine-readable text. Commercial and open-source OCR systems (like Abbyy, OCRopus, Tesseract etc.) have traditionally been optimized for contemporary documents like books, letters, memos, and other end-user documents. However, these systems are difficult to use equally well for digitizing historical document images, which contain degradations like non-uniform shading, bleed-through, and irregular layout; such degradations usually do not exist in contemporary document images.
The open-source anyOCR is an end-to-end OCR pipeline, which contains state-of-the-art techniques that are required for digitizing degraded historical archives with high accuracy. However, high accuracy comes at a cost of high computational complexity that results in 1) long runtime that limits digitization of big collection of historical archives and 2) high energy consumption that is the most critical limiting factor for portable devices with constrained energy budget. Therefore, we are targeting energy efficient and high throughput acceleration of the anyOCR pipeline. Generalpurpose computing platforms fail to meet these requirements that makes custom hardware design mandatory. In this paper, we are presenting a new concept named iDocChip. It is a portable hybrid hardware-software FPGA-based accelerator that is characterized by low footprint meaning small size, high power efficiency that will allow using it in portable devices, and high throughput that will make it possible to process big collection of historical archives in real time without effecting the accuracy.
In this paper, we focus on binarization, which is the second most critical step in the anyOCR pipeline after text-line recognizer that we have already presented in our previous publication [21]. The anyOCR system makes use of a Percentile Based Binarization method that is suitable for overcoming degradations like non-uniform shading and bleed-through. To the best of our knowledge, we propose the first hardware architecture of the PBB technique. Based on the new architecture, we present a hybrid hardware-software FPGA-based accelerator that outperforms the existing anyOCR software implementation running on i7-4790T in terms of runtime by factor of 21, while achieving energy efficiency of 10 Images/J that is higher than that achieved by low power embedded processors with negligible loss of recognition accuracy.
- {n. d.}. ABBYY. https://www.abbyy.com/en-eu/. ({n. d.}).Google Scholar
- {n. d.}. Digital Multimeter Vilcraft. http://www.produktinfo.conrad.com/datenblaetter/100000-124999/124608-an-01-ml-VOLTCRAFT_VC_870_DMM__K__de_en_fr_nl.pdf. ({n. d.}).Google Scholar
- {n. d.}. Efficiently Implementing Dilate and Erode Image Functions - Stephen Ostermiller. https://blog.ostermiller.org/dilate-and-erode. ({n. d.}).Google Scholar
- {n. d.}. Narrenschif. http://kallimachos.de/kallimachos/index.php/Narragonien. ({n. d.}).Google Scholar
- {n. d.}. OCRopus. https://github.com/tmbdev/ocropy. ({n. d.}).Google Scholar
- {n. d.}. Omnipage. www.nuance.de/for-individuals/by-product/omnipage/index. htmwww.nuance.de/for-individuals/by-product/omnipage/index.htm. ({n. d.}).Google Scholar
- {n. d.}. Tesseract. https://github.com/tesseract-ocr. ({n. d.}).Google Scholar
- M Z Afzal, M. Kramer, Syed Saqib Bukhari, M R Yousefi, Faisal Shafait, and T M Breuel. 2014. Robust Binarization of Stereo and Monocular Document Images Using Percentile Filter. Vol. 1. Springer, 139--149. Google ScholarDigital Library
- Luis Alvarez and Luis Mazorra. 1994. Signal and Image Restoration Using Shock Filters and Anisotropic Diffusion. SIAM J. Numer. Anal. 31, 2 (April 1994), 590--605. Google ScholarDigital Library
- Thomas M Breuel, Adnan Ul-Hasan, Mayce Al-Azawi, and Faisal Shafait. 2013. High-performance OCR for printed English and Fraktur using LSTM networks. In Document Analysis and Recognition (ICDAR), 2013 12th International Conference on. IEEE, 683--687. Google ScholarDigital Library
- Syed Saqib Bukhari, Ahmad Kadi, Mohammad Ayman Jouneh, Fahim Mahmood Mir, and Andreas Dengel. 2017. anyOCR: An Open-Source OCR System for Historical Archives. The 14th IAPR International Conference on Document Analysis and Recognition (ICDAR2017), Kyoto, Japan (2017).Google ScholarCross Ref
- Alex Graves, Santiago Fernández, Faustino Gomez, and Jürgen Schmidhuber. 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. In Proceedings of the 23rd international conference on Machine learning. ACM, 369--376. Google ScholarDigital Library
- J. He, Q. D. M. Do, A. C. Downton, and J. H. Kim. 2005. A comparison of binarization methods for historical archive documents. In Eighth International Conference on Document Analysis and Recognition (ICDAR'05). 538--542 Vol. 1. Google ScholarDigital Library
- E. Kavallieratou and S. Stathis. 2006. Adaptive Binarization of Historical Document Images. In 18th International Conference on Pattern Recognition (ICPR'06), Vol. 3. 742--745. Google ScholarDigital Library
- Kazuya Kawakami. 2008. Supervised Sequence Labelling with Recurrent Neural Networks. Ph.D. Dissertation. Ph. D. thesis, Technical University of Munich.Google Scholar
- F. Kheiri, S. Samavi, and N. Karimi. 2017. Hardware design for binarization and thinning of fingerprint images. ArXiv e-prints (Oct. 2017). arXiv:cs.CV/1710.05749Google Scholar
- Claudie Faure Nicole Vincent Khurram Khurshid, Imran Siddiqi. 2009. Comparison of Niblack inspired binarization methods for ancient documents. (2009), 7247 - 7247 - 9 pages.Google Scholar
- M. H. Najafi and M. E. Salehi. 2016. A Fast Fault-Tolerant Architecture for Sauvola Local Image Thresholding Algorithm Using Stochastic Computing. IEEE Transactions on Very Large Scale Integration (VLSI) Systems 24, 2 (Feb 2016), 808--812.Google ScholarDigital Library
- N. Otsu. 1979. A Threshold Selection Method from Gray-Level Histograms. IEEE Transactions on Systems, Man, and Cybernetics 9, 1 (Jan 1979), 62--66.Google ScholarCross Ref
- Puneet and Naresh Garg. 2013. Binarization Techniques used for Grey Scale Images. International Journal of Computer Applications 71, 1 (June 2013), 8--11.Google Scholar
- Vladimir Rybalkin, Norbert Wehn, Mohammad Reza Yousefi, and Didier Stricker. 2017. Hardware architecture of bidirectional long short-term memory neural network for optical character recognition. In Proceedings of the Conference on Design, Automation & Test in Europe. European Design and Automation Association, 1394--1399. Google ScholarDigital Library
- J. Sauvola, T. Seppanen, S. Haapakoski, and M. Pietikainen. 1997. Adaptive document binarization. In Proceedings of the Fourth International Conference on Document Analysis and Recognition, Vol. 1. 147--152 vol.1. Google ScholarDigital Library
- Brij Mohan Singh, Rahul Sharma, Ankush Mittal, and Debashish Ghosh. 2011. Parallel Implementation of Otsu's Binarization Approach on GPU. International Journal of Computer Applications 32, 2 (October 2011), 16--21.Google Scholar
- Brij Mohan Singh, Rahul Sharma, Ankush Mittal, and Debashish Ghosh. 2011. Parallel Implementation of Souvola's Binarization Approach on GPU. International Journal of Computer Applications 32, 2 (October 2011), 28--33.Google Scholar
- Robert A. Wagner and Michael J. Fischer. 1974. The String-to-String Correction Problem. J. ACM 21, 1 (Jan. 1974), 168--173. Google ScholarDigital Library
- Jeng-Daw Yang, Yung-Sheng Chen, and Wen-Hsing Hsu. 1994. Adaptive Thresholding Algorithm and Its Hardware Implementation. Pattern Recogn. Lett. 15, 2 (Feb. 1994), 141--150. Google ScholarDigital Library
- Mohammad Reza Yousefi, Mohammad Reza Soheili, Thomas M Breuel, Ehsanollah Kabir, and Didier Stricker. 2015. Binarization-free ocr for historical documents using lstm networks. In Document Analysis and Recognition (ICDAR), 2015 13th International Conference on. IEEE, 1121--1125. Google ScholarDigital Library
- Mohammad Reza Yousefi, Mohammad Reza Soheili, Thomas M Breuel, and Didier Stricker. 2015. A comparisonof 1D and 2D LSTM architectures for the recognition of handwritten Arabic. In Document Recognition and Retrieval XXII, Vol. 9402. International Society for Optics and Photonics, 94020H.Google Scholar
Index Terms
- iDocChip: A Configurable Hardware Architecture for Historical Document Image Processing: Percentile Based Binarization
Recommendations
Efficient Hardware Architectures for 1D- and MD-LSTM Networks
AbstractRecurrent Neural Networks, in particular One-dimensional and Multidimensional Long Short-Term Memory (1D-LSTM and MD-LSTM) have achieved state-of-the-art classification accuracy in many applications such as machine translation, image caption ...
Rapid Implementation of Embedded Systems using Xilinx Zynq Platform
SEEDA-CECNSM '16: Proceedings of the SouthEast European Design Automation, Computer Engineering, Computer Networks and Social Media ConferenceIn any digital system design, it is crucial to achieve the lowest time-to-market possible. Indeed, that need has pushed large FPGA manufacturers to produce SoCs which will implement reprogrammable logic along with CPU and DSP cores. Especially, during ...
When Massive GPU Parallelism Ain’t Enough: A Novel Hardware Architecture of 2D-LSTM Neural Network
Multidimensional Long Short-Term Memory (MD-LSTM) neural network is an extension of one-dimensional LSTM for data with more than one dimension. MD-LSTM achieves state-of-the-art results in various applications, including handwritten text recognition, ...
Comments