In the article features of functioning of the most common optical character recognition (OCR) tools EasyOCR and TesserOCR are considered; experimental analysis of results of work of these OCR is given for the most widespread data sources, such as electronic text document, internet resource, and banner; based on analysis of the experiment results from the comparative analysis of considered OCRs by time and accuracy was made; effective algorithm of using an OCR and recommendations for their application was offered for not-distorted data, for slightly and highly distorted data.
Optical character recognition (OCR) is widely used for extracting text information from printed
documents, books, posters, business cards, electronic text documents, and internet resources, and also for
automation of data entry [
The output of recognition is a computerized text that can be easily handled [
There are several approaches to recognize text data from images. Core algorithms of OCR systems
belong to one of two basic types: matrix matching and feature extraction with different efficiency in
different situations [
Matrix matching (pattern matching, pattern recognition, image correlation) means comparing an image with stored glyph pixel-by-pixel. Feature extraction decomposes glyphs into features, that comparing with a vector-based representation of the character. More often neural networks (NN) are used for the detection of features [10, 11]. This approach more accurate than matrix matching, so most modern OCR systems using it. Regarding using NNs in the second approach, even one algorithm could show different results depending on hyperparameters, image quality, etc. So, sometimes it's a quite serious challenge to choose the appropriate OCR for a specific data source.
The purpose of this article – to do a comparative analysis of efficiency of the most widespread OCR (EasyOCR and TesserOCR) based on results of text recognition experiments in different conditions of data acquisition; to describe features of using considered OCRs for common data sources; to offer an algorithm and recommendations for effective practical application of EasyOCR and TesserOCR for standard conditions of use. Therefore, the experiments are focused on assessing the effectiveness of the state of the art OCR application on standard computers without special support of GPU and HPC computing.
The issues of image preprocessing are not considered, since this is a rather broad topic that requires additional research. The images are considered to have been prepared and no additional preprocessing steps are taken, besides than those OCR systems perform by default.
degree of distortion were taken
All artifacts were applied to photos using corresponding command-line keys for TRDG on the dataset generation stage. A list of keys can be found in the user manual of the TRDG utility [22].
All data, used in our work, and also results of recognition and summary table of results are in the data warehouse [23]. Next UML-diagram (Figure 1) visualizes the file structure of the using dataset.
Both services TesserOCR and EasyOCR were tested on the same input data and hardware under one test script, measuring text recognition time for both services.
To ensure a level playing field for both neural networks, based on which the tested services work, we are not accelerating them explicatively by using such common practices as batching or using multithreading; was left default settings, given by developers. Both NNs have flexible API and wide possibilities for customization, therefore, with correctly selected hyperparameters during neural network training for a specific situation, any of them may show much better results, especially in case of using GPU and HPC, but then it will be difficult to objectively compare their productivity.
Since networks accept files in different ways, this difference was also taken into consideration. EasyOCR can accept directly most image file formats, TesserOCR by default accept only a few file formats directly when transmitting their addresses on a drive, which makes it less flexible than EasyOCR.
TesserOCR relies in work on using PIL (Python Imaging Library) for transmitting images. Actually, PIL just uploads images from drive into intermediate format. For these purposes, it can also be used OpenCV, scikit, etc. The work of two libraries in conjunction is described in the user's manual and examples from developers, which can be found on the official project page. Preliminary experiments during the phase of search of researched libraries show, that image transmitting by using PIL does not significantly affect text recognition speed and quality.
Based on this, and also on the fact that this approach is actually standard for TesserOCR, it was used in NNs testing, to put them on a level playing field. For the purity of the experiment, for TesserOCR was also counted time for using PIL functions, the cause of EasyOCR should also perform image transformation in a recognition method call.
During experiment performance, test script one by one launches both NNs for recognition of every image from the dataset, marks the start and end time of text recognition, and saves time value and recognized data for each example in JSON format for further processing. For experiment purity, tests were run 3 times to reduce the possible impact of low-level processes of OS or the behavior of the hardware part of the PC. In total it is 122 recognitions per one iteration (61 per each NN), and a total of 366 recognitions was made during the experiment.
The experiment has been run on a laptop HP Elitebook 8460p with Mobile DualCore Intel Core i52540M CPU, frequency 3100 MHz (31 x 100) with 4 threads ant 4 Gb RAM DDR3 and Intel Sandy Bridge-MB – Integrated Graphics Controller (MB GT2 1.3GHz+) under the manage of OS Kali Linux. This computer was taken as an instance of the target device class of the experiment.
Estimates of the results of experiments are presented in Table 1 – Table 12.
Table 1 shows recognition time for the scanned fragment of a book page in seconds, where full / half/quarter in the file name – page filling degree, high / medium / worst – image quality level.
Table 12 shows the probability of correct character recognition on clear photos during a series of experiments.
We have found that TesserOCR much faster than EasyOCR in all test cases. Also, TesserOCR shows quite better accuracy in most cases, but for the recognition of small hard distorted images, EasyOCR shows better accuracy. So, for images of average big plain text, such as book pages the best OCR is TesserOCR, but for short texts with a lot of noise speeded up EasyOCR may be preferred.
Was found a direct relationship between the accuracy of recognition within one format and a time of recognition, and also direct dependence on the quality of recognizable image, regardless of file format. With lower quality, recognition is faster, at the same time, the inverse dependency of the recognition quality on the image quality is weakly expressed and does not always manifest itself. With different image formats, the BMP format often falls out of the general pattern, and corresponds to the average quality of the JPG format, slightly inferior to it in time, but superior in quality. In fact, the file format does not directly affect the recognition quality, since images are usually converted to an intermediate format, but the quality of the image stored on the disk directly affects the recognition quality.
During the control experiments, in the same conditions, and on the same inputs, was detected spread of time values between tests, most likely due not to functioning features of computer hardware and software, but due to features of algorithms of ОCR. Regardless of differences of received results, results of al a grouped around mean, the spread relative to which is the greater, the higher the quality, and the larger the volume of the recognized text. For each test on EasyOCR, the spread is much larger than for TesserOCR.
TesserOCR is much faster than EasyOCR and better recognize regular book text, but if natural scene images are taken, then EasyOCR works better. Of course, EasyOCR is still inferior to the opponent in time, but not in quality, recognizing text fully or partially, where TesserOCR doesn’t recognize anythin In Table 13 given superiority rates for TesserOCR and EasyOCR by recognition time and quality and symbols per second ratio. Constructing this table, we calculated average indicators of time and accuracy by given according to the tables above and divided corresponding indicators for EasyOCR into indicators for TeserOCR. Their quality (accuracy) is expressed by the ratio of correct recognition probability. Our calculations may be expressed by formula where R – superiority rate for corresponding indicators; n – numbers of tests; e – indicators, obtained in experiments for EasyOCR; t – indicators, obtained in experiments for TesserOCR. Figures 2 and 3 provide a visualized presentation of obtained superiority rates. Figure 4 visualizes the superiority rate for TesserOCR and EasyOCR by symbols per second ratio.
Analyzing obtained experimental results it is possible to formulate algorithms and recommendations for practical use of reviewed text recognition tools. The common algorithm of text recognition consists of 2 stages: 1) preprocessing and 2) recognition of prepared text. In the preprocessing stage, it is advisable to convert images into bitmaps before submitting them to the neural network, e.g. using PIL (Python Imaging Library) or any other appropriate image processing library (OpenCV, scikit, etc), if the image was originally not in format BMP. It will give the most balanced ratio of time to quality among other formats and also will provide more effective work with images. Converting into bitmap gives unification of image processing code in the final application, besides, in some situations, bitmap capture from a screen is the fastest way of image feed. Nevertheless, if the priority parameter of recognition is speed, then it is advisable to convert images into lighter formats with lower quality. In this case, the quality of recognition without preliminary retraining or hyperparameter tuning will dropdown. For effective use of both frameworks, it is necessary to do hyperparameter tuning, to train a neural network for specific input data types, and also to convert images into bitmaps before recognition. In this case, performance will be much better, than the one described as a result of the experiment. According to experiment results, the TesserOCR library can be recommended for recognition of scanned books, screenshots, soft blurred and clear photos. While the EasyOCR library can be recommended for recognition of advertisements, banners, and hard distorted photos (i.e. natural scene images). Based on the results of the analysis it should be noted that nowadays the EasyOCR library is actively developed and in the coming years by certain parameters can surpass TesserOCR, especially for the conditions of neural network optimization and overclocking.
In the work was set up the experiment and made a comparative analysis of the performance of the most common OCRs (TesserOCR and EasyOCR) by time and by the accuracy of recognition in equals conditions (with default configuration). Described application features of both services for different data sources, which can appear during the real use of neural networks. Received results describe only neural network performance without training stage and give only a general understanding of the functioning of each of them and about their field of application at the time of writing of the article. A complete analysis of the capabilities of each NN is needed deeper researches, which, however, was not the aim of this work. They imply serious modification and optimization of the code, obtained just after installation, programming on the lower level, retraining of NN, provided by developers or image preprocessing. According to the analysis of the experimental results, proposed algorithms and recommendations for effective practical application of reviewed OCRs. Besides, it is also recommended to do configuration and training of NN for each specific case (data source) for obtaining the best performance.
It is also important to remember that the results of the experiments carried out are relevant only to standard computer without the support of GPU and HPC calculations. This is done specifically to evaluate the performance of real applications that will run on the same standard computers.
[7] An Introduction to OCR for Beginners, 2020. URL: https://towardsdatascience.com/an-introductionto-optical-character-recognition-for-beginners-14268c99d60. [8] What is Cognitive Computing? How are Enterprises benefitting from Cognitive Technology, 2019.
URL: https://towardsdatascience.com/what-is-cognitive-computing-how-are-enterprises-benefittingfrom-cognitive-technology-6441d0c9067b. [9] Kyrychenko I., Gruzdo I., Tereshchenko G., Cherednichenko O. Applıcatıon of Paragraphs Vectors Model for Semantıc Text Analysıs, in: Proceedings of the 4th International Conference on Computational Linguistics and In-telligent Systems (COLINS-2020), Lviv, Ukraine, April 23-24, 2020. – Volume I, pp. 283-293. [10] Ian Goodfellow, Yoshua Bengio, Aaron Courville Deep Learning, 3rd. ed., MIT Press, 2016, 787p.
DOI/ISBN: 0262035618. [11] Peter Norvig , Stuart Russell Artificial Intelligence: A Modern Approach, 3rd. ed., Pearson
Education Limited, 2016, 1152p. DOI/ISBN: 9780136042594. [12] Kyrychenko I., Proniuk G., Geseleva N., Tereshchenko G. Spatial interpretation of the notion of relation and its application in the sys-tem of artificial intelligence, in: Proceedings of the 3rd International Conference on Computational Linguistics and In-telligent Systems (COLINS-2019), Kharkiv, Ukraine, April 18-19, 2019. – pp. 266-276. [13] Rafael C. Gonzalez, Richard E. Woods Digital Image Processing, 4th. ed., Pearson/Prentice Hall, 2018, 1168p. DOI/ISBN: 9780133356724. [14] Chupryna A., Smelyakov K., Hvozdiev M., Sandrkin D. Gradational Correction Models Efficiency Analysis of Low-Light Digital Image, in: Proceedings of the 2019 Open Conference of Electrical, Electronic and Information Sciences (eStream), 25 April 2019, Vilnius, Lithuania. – P. 1-6. doi: 10.1109/eStream.2019.8732174. [15] Datsenko A., Skrypka V., Akhundov A., Smelyakov K. Efficiency of Image Reduction Algorithms with Small-Sized and Linear Details, in: Proceedings of the 2019 IEEE International ScientificPractical Conference Problems of Infocommunications, Science and Technology (PIC S&T), 8-11 Oct. 2019, Kyiv, Ukraine. – P. 745-750. doi: 10.1109/PICST47496.2019.9061250. [16] Sharonova, N., Kanishcheva, O., Image and video tag aggregation // CEUR Workshop Proceedings,
Vol. 2268, 2017, pp. 161-172, 2017. [17] Kanishcheva, O.. Cherednichenko, O., Sharonova, N. Image Tag Core Generation // Proceedings of the 1st International Workshop on n Digital Content & Smart Multimedia (DCSMart 2019).
Lviv, Ukraine, December 23-25, 2019, pp. 35-44. [18] Cherednichenko, O., Yanholenko, O., Vovk, M., Sharonova, N. Towards structuring of electronic marketplaces contents: Items normalization technology // CEUR Workshop Proceedings. Vol. 2604, 2020, P. 44-55. [19] Smelyakov K., Yeremenko D., Sakhon A., Polezhai V., Chupryna A. Braille character recognition based on neural networks, in: Proceedings of the IEEE Second International Conference on Data Stream Mining & Processing (DSMP), August 21-25. – 2018. – P. 509-513. [20] Tesserocr, 2020. URL: https://github.com/sirfz/tesserocr. [21] EasyOCR, 2021. URL: https://www.jaided.ai/easyocr. [22] TextRecognitionDataGenerator, 2020. URL: https://github.com/Belval/TextRecognitionDataGenerator. [23] Data storage with our data and results, 2021. URL: https://drive.google.com/drive/folders/1dy2gEDBf38t_TY1933g57JyDJ2J4MmWA?usp=sharing.