MIDV-2020: A COMPREHENSIVE BENCHMARK DATASET FOR IDENTITY DOCUMENT ANALYSIS Текст научной статьи по специальности «Медицинские технологии»

MIDV-2020: a comprehensive benchmark dataset for identity document analysis

K.B. Bulatov1-2, E.V. Emelianova2-3, D.V. Tropin1-2-4, N.S. Skoryukina1-2, Y.S. Chernyshova1-2,

A.V. Sheshkus1-2, S.A. Usilin1-2, Z. Ming 5, J.-C. Burie 5, M. M. Luqman 5, V.V. Arlazarov12 1 Federal Research Center «Computer Science and Control» or Russian Academy of Sciences, Moscow, Russia;

2 Smart Engines Service LLC, Moscow, Russia;

3 National University of Science and Technology «MISiS», Moscow, Russia;

4 Moscow Institute of Physics and Technology (State University), Moscow, Russia;

5 L3i Laboratory, La Rochelle University, La Rochelle, France

Abstract

Identity documents recognition is an important sub-field of document analysis, which deals with tasks of robust document detection, type identification, text fields recognition, as well as identity fraud prevention and document authenticity validation given photos, scans, or video frames of an identity document capture. Significant amount of research has been published on this topic in recent years, however a chief difficulty for such research is scarcity of datasets, due to the subject matter being protected by security requirements. A few datasets of identity documents which are available lack diversity of document types, capturing conditions, or variability of document field values. In this paper, we present a dataset MIDV-2020 which consists of 1000 video clips, 2000 scanned images, and 1000 photos of 1000 unique mock identity documents, each with unique text field values and unique artificially generated faces, with rich annotation. The dataset contains 72409 annotated images in total, making it the largest publicly available identity document dataset to the date of publication. We describe the structure of the dataset, its content and annotations, and present baseline experimental results to serve as a basis for future research. For the task of document location and identification content-independent, feature-based, and semantic segmentation-based methods were evaluated. For the task of document text field recognition, the Tesseract system was evaluated on field and character levels with grouping by field alphabets and document types. For the task of face detection, the performance of Multi Task Cascaded Convolutional Neural Networks-based method was evaluated separately for different types of image input modes. The baseline evaluations show that the existing methods of identity document analysis have a lot of room for improvement given modern challenges. We believe that the proposed dataset will prove invaluable for advancement of the field of document analysis and recognition.

Keywords: document analysis, document recognition, identity documents, open data, video recognition, document location, text recognition, face detection.

Citation: Bulatov KB, Emelianova EV, Tropin DV, Skoryukina NS, Chernyshova YS, Sheshkus AV, Usilin SA, Ming Z, Burie JC, Luqman MM, Arlazarov VV. MIDV-2020: a comprehensive benchmark dataset for identity document analysis. Computer Optics 2022; 46(2): 252-270. DOI: 10.18287/2412-6179-ŒÎ-1006.

Acknowledgements: This work is partially supported by Russian Foundation for Basic Research (projects 19-29-09066 and 19-29-09092). All source images for MIDV-2020 dataset were obtained from Wikimedia Commons. Author attributions for each source images are listed in the original MIDV-500 description table (ftp://smartengines.com/midv-500/documents.pdf). Face images by Generated Photos (https://generated.photos).

Introduction

Document analysis and recognition is a vast and growing research field, with studies covering topics of image processing, general computer vision, machine learning, intelligent systems, computational optics, and much more. Application of automated document analysis systems to a wide range of industries and processes created a need for more robust and precise methods of document identification and recognition, and thus an active demand for research in these fields.

A particular variety of document types which play a crucial role are identity documents, such as ID cards,

passports, driving licences, and others. These documents are usually issued by governments, have strict design and security features, and their main goal is to define, verify and prove the holder's identity. The scope of usage of automatic system for identity document analysis include simplification and automatization of data entry when filling official forms [1], remote person identification [2], remote age checking [3], Know Your Customer / Anti Money Laundering (KYC / AML) procedures [4], and provision of governmental, financial, and other services.

The global demand for reliable automated identity document processing systems lead to the need of developing new methods for document detection, identifica-

tion, and location; segmenting the document into relevant fields; extraction of text fields and graphical elements, such as signatures and faces; as well as anti-spoofing, validating the document authenticity, and detection of markers of potential fraud [5]. The variation of input sources has also been growing: while traditional methods of scanning identity documents using flatbed or specialized scanners are still relevant, with the accompanying problems of scale, image definition, and image coloration being unknown or uncontrolled, remote identification procedures now require the document analysis systems to process camera-captured photos, as well as perform ID recognition in a video stream. Thus, modern identity document analysis systems which are capable of processing photo and video data must address such difficulties as unknown pose, capture angle and orientation, possible occlusions, uncontrolled lighting, shadows and highlights [6]. Since the processing of ID copies uploaded by the customers within a KYC process should support both scans and photos [5], the identity document analysis systems employed for such tasks are required to be robust not only to the specifics of a single input source, but to a mixed stream of data from different input sources.

One of the most important aspects of effective research of new document analysis methods is the availability of openly accessible datasets for training and benchmarking. However, when it comes to identity documents, this issue becomes very hard to resolve, since identity documents by their nature contain sensitive personal data, which is protected by worldwide regulations.

As was mentioned in the previous sections, since identity documents by their nature contain sensitive information, there are very few publicly available datasets of identity document images, and those which exist contain either partial information, or contain synthetic examples of ungenuine documents. Existing datasets dedicated specifically to identity document images include LRDE Identity Document Image Database (LRDE IDID) [7], the recently published Brazilian Identity Document Dataset (BID Dataset) [4], and the Mobile Identity Document Video dataset family (MIDV) [8, 9], to which the dataset presented in this paper also belongs. Some larger datasets, dedicated to address the issues of a broader document analysis problem, such as the ones from SmartDoc family [10], also contain identity document images.

Existing datasets of identity document images have disadvantages which present themselves when the datasets are used by researchers as benchmarks for specific identity document analysis tasks. The LRDE IDID [7] and the identity documents subset of SmartDoc [10] comprise a small amount of document samples, which allows them only to be used as reference benchmarks, without deeper analysis of identity processing methods. BID Dataset [4] addresses that issue, featuring 28800 synthetically generated document images with 8 different document types. At the same time, the images of BID Dataset were generated with artificial inscription of text

field values over the automatically masked document regions, which might lead to the field presentation widely different from the one on actual documents, and the document owner faces were blurred in each document image, which make the dataset impossible to use for evaluation of face detection and location methods. In addition, BID Dataset only features ideally cropped documents, however some further datasets from the same team include synthetic scans and photographs as well.

The first dataset of the MIDV family was MIDV-500 [8], which contained 500 video clips of 50 identity documents, 10 clips per document type. The identity documents had different types, and mostly were «sample» or «specimen» documents which could be found in Wiki-Media and which were distributed under public copyright licenses. The dataset focused on mobile video capture, featured clips shot using two smartphones (Apple iPhone 5 and Samsung Galaxy S3) under five distinct conditions: «Table», «Keyboard», «Hand», «Partial», and «Clutter».

The conditions represented in MIDV-500 thus had some diversity regarding the background and the positioning of the document in relation to the capturing process, however they did not include variation in lighting conditions, or significant projective distortions. To address the latter issues a dataset MIDV-2019 [9] was later published as an extension of MIDV-500, which featured video clips captured with very low lighting conditions and with higher projective distortions. Video clips of MIDV-2019 were captured using Apple iPhone XS Max and Samsung Galaxy S10. The dataset was also supplemented with photos and scanned images of the same document types [11] to represent typical input for server-side identity document analysis systems.

Since its publication, MIDV-500 dataset and its extension MIDV-2019 were used to evaluate the methods of identity document images classification [12 - 14]; identity document location [11, 15], including the methods based on semantic segmentation [16]; detecting of faces on images of identity documents [17]; and methods related to text fields recognition, including single text line recognition [18], per-frame recognition results combination [19, 20] and making a stopping decision in a video stream [21, 22]. The dataset was also used to evaluate the methods of choosing a single best frame in the identity document video capture [23] and assessing the quality of the frame for its processing by an identity analysis system [24], detection and masking of sensitive and private information [25] and general ID verification [26].

The main disadvantage of the datasets of the MIDV family is the scarcity of different document samples — all images were made using the same 50 physical document samples. The fact that the text field data and the graphical data (signatures, faces) are not variable, makes it harder to produce reliable evaluations of identity document analysis algorithms, or even to calculate meaningful benchmarks [6].

In this paper, we present a new dataset MIDV-2020, based on the 10 document types featured previously in MIDV-500 and MIDV-2019. One of the main goals of the MIDV-2020 dataset, presented in this paper, is to provide variability of the text fields, faces, and signatures, while retaining the realism of the dataset. The dataset consists of 1000 different physical documents (100 documents per type), all with unique artificially generated faces, signatures, and text fields data. Each physical document was photographed, scanned, and for each a video clip was captured using a smartphone. The ground truth includes ideal text field values, geometrical position of documents and faces in each photo, scan, and video clip frame (with 10 frames-per-second annotation). We believe that MIDV-2020 will prove invaluable for the advancement of the field of document analysis and recognition, as well as serve as a benchmark for modern identity document processing systems.

1. Dataset

In this section, we will present the basic composition of the dataset, description of its contents and annotations.

1.1. Composition of the dataset

The set of base document types for MIDV-2020 comprises 10 document types, each present in previously published MIDV-500 and MIDV-2019 datasets. The identity document types of MIDV-2020 are listed in tab. 1, with the codes of the PRADO database [27] for each document type except the Internal passport of Russia. 100 sample documents were created for each of the 10 document types present in the dataset.

Tab. 1. Document types featured in MIDV-2020

№ Document type code Description PRADO code MIDV- 500 code

1 alb id ID Card of Albania ALB-B0-01001 01

2 aze_passport Passport of Azerbaijan AZE-A0-02002 05

3 esp id ID Card of Spain ESP-B0-03001 21

4 est id ID Card of Estonia EST-B0-03001 22

5 fin id ID Card of Finland FIN-B0-06001 24

6 grc passport Passport of Greece GRC-A0-03003 25

7 lva passport Passport of Latvia LVA-A0-01004 32

8 rus_internal-passport Internal passport of Russia n/a 39

9 srb passport Passport of Serbia SRB-A0-01001 41

10 svk_id ID Card of Slovakia SVK-B0-05001 42

To create unique sample documents the original sample images obtained from Wikimedia Commons (we used the same source images as in MIDV-500 [8]) were edited to remove non-persistent data (such as signature, photo, and text field values). Then, the generated text field values for each field were added to the images using a font which closely resembles the original font of the sample. In addition, for each document a mock signature was drawn, vaguely resembling the spelling of the last name, and a unique face picture was added.

The values of gender, birth date, issue date, and expiry date were generated in accordance with the specifics of issuing countries and a pre-set distribution of age and gender parameters:

1. 80 % of generated documents correspond to adult holders (ages 18 through 60), 10 % of generated documents correspond to seniors (ages 60 through 80) and 10 % to children and adolescents (17 or less) depending on the minimum age for the document issue.

2. 50 % of the generated documents correspond to female holders, and 50 % to male holders.

Names and addresses were generated using the databases of existing names and examples of addresses available online, using Wikipedia listings of names [28], and online name generators [29].

Artificially generated face pictures by the Generated Photos service [30] were used to provide a unique face for each document. The service lists StyleGAN [31] as the approach used to create artificial face images. The images were taken either in color or in grayscale, depending on the original document sample, and were repeated if the document contained multiple copies of the face picture with opacity corresponding to the original sample. The pictures were selected to approximately match the holder age.

The template images prepared in such a way were printed to scale on a glossy photo paper, laminated, and cropped. All four corners of the card-sized documents, and the bottom two corners of the passport-sized documents were rounded using a 4 mm corner rounder.

The original template images, which were used for printing, along with their annotation files are placed in the dataset within the «templates.tar» archive. There are 1000 template images in total. The images are divided according to the document type and for each document type they are numbered from «00.jpg» to «99.jpg». The contents of the «templates.tar» is depicted in fig. 1 of the Appendix. An example of a template image is presented in fig. 1.

REPUBUKA E SHQIPERISE ...................

mbiemri/surname

Aga ni

emri/given name

Adnan

shtetësia/nationality

Shqiptare/Albanian

vendlindja/piace of birth

Bulçar, ALB

datelir ^»a/date of birth

20-11-1991

data e teshinçiit/date of issue

13-04-2019

autoriteti lëshues/authority

MB

nr. letërnjoftim/card ne

367253746

gjinia/sex

M

data e skadimit/date of expiry

13-04-2029

nr. personal/personal no.

J11120296E

J

Fig. 1. An example of a template image (image albid / 00)

For each image a corresponding annotation file in JSON format is provided, readable with VGG Image An-notator v2.0.11 [32], available on its website [33] (see the

fragment of the annotation tool screenshot in fig. 2). The annotation for each image includes rectangular bounding boxes of the main document holder photo (named «photo»), the bounding box of the face oval (named «face»), the bounding box of a signature field (named «signature») and the rectangles corresponding to the positions of text fields. For each text field its exact value is provided. The upper and lower boundaries of the text field rectangle correspond to the cap line and baseline respectively, and for each text field additional information is provided to indicate whether the field has lowercase letters, descenders, or ascenders. Finally, for document types where some text fields are oriented vertically, an additional orientation attribute is provided for such text fields which indicates the angle of the field's counterclockwise rotation expressed in degrees. The format of the template image annotations is given in fig. 5 of the Appendix.

REPUBLIKA E SHQIPËRISË BAN^

Fig. 2. An example of a template image with fields annotation (image alb id/ 00)

1.2. Scans

Each physical document sample was scanned using Canon LiDE 220 and Canon LiDE 300 scanners, in two different modes. The first mode shows the document in an upright position, with a slight margin, near the top-right corner of the scanned image. The margin is enforced with a pink piece of paper. The second mode shows the document in an arbitrary place within the scanned image frame, and rotated to an arbitrary angle. All scanned images have the resolution of 2480 x 3507 pixels, they have not been automatically cropped. Examples of an «upright» and «rotated» scans are presented in fig. 3.

The scanned images were captured in TIFF format, then converted to JPEG using ImageMagick 7.0.11 with default parameters. The JPEG images with their annotations are placed in the dataset within the archives «scan_upright.tar» and «scan_rotated.tar». The original TIFF images are placed within the archives «scan_upright_tif.tar» and «scan_rotated_tif.tar». There are 1000 upright scans and 1000 rotated scans in total. The names of the scanned images correspond to the names of the template image from which the physical document was created.

The contents of the annotated scans archives «scan_upright.tar» and «scan_rotated.tar» follow the

same structure, depicted on fig. 1 of the Appendix. The contents of the archives with original TIFF images obtained from scanners («scan_upright_tif.tar» and «scan_rotated_tif.tar») follow a similar structure, but without the annotations (see fig. 2 of the Appendix).

Fig. 3. Examples of MIDV-2020 scans: (a) upright scan,

(b) rotated scan (image alb_id/00). Due to wide margins

of the original scans, the images presented in the figure are cropped with centering on the document

The annotations for scanned images are provided in JSON format readable with VGG Image Annotator v2.0.11 and feature the bounding boxes of the face oval (marked with a field name «face») and the document quadrangle (marked with a field name «doc_quad»). The first vertex of the quadrangle always corresponds to the top-left corner of the physical document, and the rest of the vertices are provided in a clockwise order. The format of the annotations is given in fig. 6 of the Appendix.

1.3. Photos

For each physical document sample a photo was taken, given various conditions and two smartphones. Half of the photos were captured using Apple iPhone XR, and the other half using Samsung S10. The following capturing conditions were used:

1. Low lighting conditions (20 documents of each type);

2. Keyboard as a background (10 documents of each type);

3. Natural lighting, captured outdoors (10 documents of each type);

4. Table as a background (10 documents of each type);

5. Cloth with various textures as a background (10 documents of each type);

6. Text document as a background (10 documents of each type);

7. High projective distortions of the document (20 documents of each type);

8. Highlight from the sun or lamp hides a portion of the document (10 documents of each type).

Examples of photos captured with different conditions are presented in fig. 4.

The ranges of codes of the document samples which correspond to each of the listed conditions are specified in tab. 2.

All photos are stored in JPEG format and have the resolution of 2268 x 4032 pixels. The images with annota-

tions are placed in the dataset within the «photo.tar» archive. The names of the images correspond to the names of the template from which the physical document was created. There are 1000 photos in total. The contents of the pho-

tos archive «photo.tar» follow the structure similar to the one of templates and scans, as depicted in Fig.1 of the Appendix. The annotations format is the same as for the scans (see fig. 6 of the Appendix).

Fig. 4. Examples of MIDV-2020photos: (a) low lighting (image albid/70); (b) Keyboard (image aze_passport/35); (c) Outdoors (image espid/40); (d) Table (image estid/55); (e) Cloth (image finid/90); f) Text background (image grc_passport /25); (g) Projective (image lva_passport/00); (h) Highlights (image rusinternalpassport/ 65)

Tab. 2. Ranges of image codes corresponding to photo capturing conditions

Capturing conditions and smartphone models Samsung S10 Apple iPhone XR

Low lighting 80 - 89 70 - 79

Keyboard in the background 35 - 39 30 - 34

Natural lighting, outdoors 45 - 49 40 - 44

Table in the background 55 - 59 50 - 54

Cloth in the background 95 - 99 90 - 94

Text documents in the background 25 - 29 20 - 24

Projective distortions 10 - 19 00 - 09

Highlight present 65 - 69 60 - 64

1.4. Video clips

For each document sample a video clip was captured, using the similar distribution of the capturing conditions as with photos. Each clip was captured vertically, in a resolution of 2160 x 3840 pixels, with 60 frames per second. There are 1000 clips in total, each with different lengths. The numbers of the document samples which

correspond to each of the capturing conditions are specified in tab. 3.

Tab. 3. Ranges of clip codes corresponding to video capturing conditions

Capturing conditions and smartphone models Samsung S10 Apple iPhone XR

Low lighting 00 - 09 10 - 19

Keyboard in the background 20 - 24 25 - 29

Natural lighting, outdoors 60 - 64 65 - 69

Table in the background 30 - 34 35 - 39

Cloth in the background 40 - 44 45 - 49

Text documents in the background 50 - 54 55 - 59

Projective distortions 70 - 79 80 - 89

Highlight present 90 - 94 95 - 99

The original captured clips were separated into frames using ffmpeg version n4.4 with default parameters. To provide a rich annotation each video clip was annotated with 10 frames per second. To this end, each 6-th frame

of each clip was retained (thus, the dataset only retained the separate frames «000001.jpg», «000007.jpg», «000013.jpg», etc.). The smallest clip has 38 frames, the largest has 129 frames. Overall, the dataset includes 68409 annotated video frames.

Frames with the corresponding annotations are placed in the dataset within the «clips.tar» archive. It's structure is demonstrated on fig. 3 of the Appendix. The annotation of each frame has the same format as for the photos and scanned images (see fig. 6 of the Appendix). The original video files are located in the «clips_video.tar» archive, which has the structure depicted in fig. 4 of the Appendix. The video clips were stripped of audio channel.

2. Benchmarks

While the main goal of the paper is to present an identity document dataset MIDV-2020, in order to provide a baseline for future research involving the dataset, in the following sections several benchmarks using MIDV-2020 will be presented. Based on the topics for which the datasets of the MIDV family were used (see the Introduction section), the benchmark will include the task of content-independent document boundaries location, semantic segmentation of the document body from the background, feature-based document location and type identification, text fields recognition, and face detection. Since all these document analysis subtasks are crucial for identity document recognition systems, the baselines provided in this section will serve not only as a representation of the existing methods' performance on the presented dataset, but also as a reference for experimental evaluation of new methods and the future development of the field.

2.1. Content-independent document boundaries location

Detection and location of a document in an image or a video frame is one of the first major steps in an identity document analysis system. When the set of possible document types is not known in advance, or if the knowledge of the general structure of the document content is not available, the systems have to resolve to content-independent methods for document location. One of the types of such methods is locating document boundaries, under the assumption that the document is a planar rectangular object with a known aspect ratio (or a known set of possible aspect ratios).

Consider an image which contains an identity document. Let us assume that the document is a planar rigid rectangle with a known aspect ratio. The corners of the rectangle could be either sharp or rounded. The content of the document is not known in advance, however we can assume that the content color is distinguishable from the background. The image is captured using a camera or a scanner. The capturing conditions are uncontrolled: the background, lighting, and document orientation can be arbitrary, the document may partially be outside the frame or occluded by another object. We assume that in the image there is only one document with a given aspect

ratio. The task is to estimate the coordinates of the vertices of the quadrangle corresponding to the document boundary, down to their convex re-enumeration. An input to the algorithm, besides the image I, comprises the size of the document template (henceforward, we will use t as a rectangle with vertices in points (0; 0), (w; 0), (w; h), (0; h), where w and h are linear sizes of the template). Based on this input the algorithm needs to estimate the quadrangle q which would satisfy a binary quality criterion L (q, t, m), where m represents a ground truth quadrangle.

Binary quality criterion L may be based on (i) a Jac-card score:

IoU(q, m) =

area(q n m) ; area(q u m) '

(1)

(ii) Jaccard score calculated in the coordinate system of a ground truth quadrangle m:

IoUgt(q, m, t ) =

area(Mq n t ) area(Mq u t ) '

(2)

where M designates a homography such that Mm = t; (iii) maximum distance between the corresponding vertices of q and m computed in the coordinate system of q:

D0(q, m, t ) = max t

t - Hmj |

P(t )

(3)

where H represent a homography such that Hq = t and P (t) is a perimeter of the template. Since the quadrangle of the document should be detected down to their convex re-enumeration of its vertices the final binary quality criterion based on (3) is a follows:

D4(q, m, t ) = minqWEe D0(q(i), m, t ),

(4)

where Q is the set of convex vertex re-enumerations {[a, b, c, d]; [b, c, d, a]; [c, d, a, b,]; [d, a, b, c,]}.

There exist three base approaches for detecting a quadrangle of the document boundaries: corners detection [34, 35], straight lines detection and analysis [36 - 39], and the analysis of salient regions [2, 40]. As shown in [37], the representatives of all three classes successfully solve the problem given a subset of an open dataset SmartDoc [41], which complies to the problem statement.

There is an openly available model from [34], which uses a convolutional neural network for document vertex detection. For the purposes of a baseline evaluation, we used a model trained using a dataset of A4-size documents (SmartDoc subset and own-collected). We also asked the authors of [37 - 39] to evaluate their algorithms using MIDV-2020. The method described in [37] uses straight lines (for generating quadrangle hypotheses) and salient regions (for comparing and ranging the hypotheses). This algorithm shows state-of-the-art results on the open dataset MIDV-500 [8] and can be used on low-end computational devices. The algorithms [37] and [39] both require the knowledge of the document's aspect ratio and the knowledge of the camera internal parameters (focal

length and position of a principal point). Although we do not know the latter from our problem statement, in experiments with [37] and [39] we used the focal length equal to 0.705 from the diagonal of the image and the position of principal point in the center of the image. The algorithm [38] as well as [37] uses straight lines and salient regions however does not require knowledge of the aspect ratio and internal parameters of the camera.

In order to estimate the accuracy given a dataset of images we will use the following statistics: the rate of correct answers for the two binary quality criteria: IoU > 0.9 (see tab. 4) and D4 < 0.02 (see tab. 5), as well as a mean Jaccard score IoUgt (see tab. 6). The link to the source code of the statistics implementation is published in [37]. To perform the numerical evaluation, ground truth document boundary quadrangles were used, which are present in MIDV-2020 for all types of images (except ideal document templates).

Tab. 4. Document location performance (percentage of images with IoU > 0.9) of content-independent methods on MIDV-2020

Method MIDV-2020 subset

Upright scans Rotated scans Photos Video frames

[34] 0.00 2.50 34.00 23.34

[37] 100.00 91.30 87.00 85.55

[38] 72.80 50.80 82.00 78.01

[39] 98.20 91.20 83.30 81.74

Tab. 5. Document location performance (percentage of images with D4 < 0.02) of content-independent methods on MIDV-2020

Method MIDV-2020 subset

Upright scans Rotated scans Photos Video frames

[34] 0.00 1.70 33.50 22.99

[37] 99.70 90.80 85.40 83.08

[38] 72.60 49.70 79.20 74.74

[39] 97.00 90.70 81.80 79.65

Tab. 6. Document location performance (mean value of IoUgt) of content-independent methods on MIDV-2020

Method MIDV-2020 subset

Upright scans Rotated scans Photos Video frames

[34] 0.2142 0.5462 0.6569 0.6144

[37] 0.9874 0.9320 0.8977 0.8997

[38] 0.7353 0.5381 0.8682 0.8644

[39] 0.9790 0.9192 0.8797 0.8791

Example outputs of the evaluated systems are presented in fig. 5. Note that on the fig. 5a the output of the [37] and [39] totally overlap, the same on the fig. 5c, but for systems [37 - 39].

The results of the system presented in [34], while being quite low, correspond to the data presented in the original paper: the mean value of IoUgt on the 5-th subset of the SmartDoc dataset (excluding the training subset) was reported to be 0.67. Lower results on the 5-th subset of SmartDoc was explained by the presence of additional objects near the target document. In the case of MIDV-2020 additional complications include (i) the presence of scanned images, where the documents are shifted to the top-right corner and have small relative

size; and (ii) the documents of MIDV-2020 have rounded corners, which negatively influence the second stage of the algorithm presented in [34], where the corners are recursively refined. An example of the rounded corners effect is presented in fig. 5c: the system of [34] successfully detected sharp corners, but incorrectly detected the rounded ones.

Fig. 5. Examples of document quadrangle detection in images

from different MIDV-2020 subsets: (a) Upright scans (image alb_id/00), (b) Rotated scans (image srb_passport/86),

(c) Photo (image srb_passport/51), (d) Video clip frame (frame alb_id/22/000217). Color denotes different methods: [34] - magenta, [37] - red, [38] - green, [39] - blue

As tab. 5 and 6 show, the metrics of the mean value of IoUgt may be insufficient for the evaluation and benchmarking of document location in context of ID recognition systems: the results of [34] on the «Rotated scans» subset is higher than that of [38], whereas the values of (IoU > 0.9) and (D4 < 0.2) for this method amount only to a few percent, being significantly lower than [38] on the same subset.

2.2. Document location via semantic segmentation

An alternative approach to the search of document boundaries is performing semantic segmentation of the image to determine which pixels correspond to the document and which to the background. Here we can assume that the document is pseudo-rigid, possibly warped quadrangle and unknown internal content. The task is to clas-

sify each pixel to either document or a background. The most natural metrics to evaluate such a task is to compute a pixel-wise Jaccard score [15].

To evaluate the semantic segmentation method, we used the published methods HU-PageScan [15] and HoughEncoder [16], both based on a U-net modification. HU-PageScan has an openly available source code and a pre-trained model, which was used for this evaluation. We asked the authors of HoughEncoder to evaluate their model with training on MIDV-500 [8] and MIDV-2019 [9]. The ground truth masks which are required for evaluation of semantic segmentation-based methods were generated using the ground truth document boundary quadrangles present in MIDV-2020 for all types of images (except ideal document templates). The results of the evaluation are presented in tab. 7.

Tab. 7. Document location performance (mean value ofloU) of semantic segmentation methods on MIDV-2020

Method MIDV-2020 subset

Upright scans Rotated scans Photos Video frames

[151 0.0941 0.0564 0.2680 0.1509

[161 0.0850 0.1688 0.6508 0.6792

The results of [15] were obtained using its source code supplied with the original paper. Examples of the semantic segmentation results on images from various MIDV-2020 subsets are presented in fig. 6.

Low results of the HoughEncoder method [16] on the «Upright scans» and «Rotated scans» subsets of MIDV-2020 can be explained by its training on MIDV-500 and MIDV-2019 datasets, which do not feature scanned images.

Since both methods are trained and inferred using full images, including document contents and backgrounds, they are not, strictly speaking, content-independent. This might sometimes lead to segmenting only specific parts of the documents, for example, in fig. 6e and 6h it can be seen that only the zones of the documents with visible text lines are classified as belonging to a document. The low baseline results for semantic segmentation-based document location methods as evaluated on MIDV-2020 (even if trained on MIDV-500 and MIDV-2019) show the need to continue the research and improve the existing methods for them to attain practical application for identity image analysis.

2.3. Feature-based document location and identification

The previous two sections considered document location methods which did not directly rely on the internal document content (like the boundary analysis methods), or which learned the differences between document content and background (like in the case of semantic segmentation) given a train dataset. Identity documents in general can be classified as semi-structured objects - their content consists of static elements of the document template and variable personal data. Within a single docu-

ment template type the exact or approximate position of the variable personal data is known. Let us consider the task of document type identification given a scanned or photographed image. Given a predefined set of identity document templates (which, in a general case, may contain hundreds or thousands of different templates) the task is to classify the image to one of the classes representing a template, or to a null class, which correspond to the absence of a known document type. The state-of-the-art approach [42] for solving such a problem is based on local features matching. The typical base features are keypoints and their descriptors, however the feature space can be expanded with line segments, quadrangles, vanishing points, etc. [11, 43].

a) i i (k) (i) i

Fig. 6. Examples of semantic segmentation results of images from MIDV-2020 dataset (the source images correspond to the ones presented in fig. 5. (a) source image; (b) result of [15]; (c) result of [16]; (d) source image; (e) result of [15]; f) result of [16]; (g) source image; (h) result of [15]; (i) result of [16]; (j) source image; (k) result of [15]; (l) result of [16]

To provide a baseline for evaluation of such methods using MIDV-2020 dataset we used the algorithm described in [44] :

2.

3.

1. For each ideal document template an ideal image is used to mark up the zones with variable personal data. Keypoints and their descriptors are extracted in an ideal document template image and in the analyzed image. The keypoints inside the variable data zones are discarded.

A set of direct matches is constructed. For each local descriptor of the analyzed image the closest descriptor from the set of all stored ideal image descriptors is found. The document types are ranged by the total number of the closest match occurrences with respect to all keypoints in the analyzed image.

4. Among the found matches the symmetric ones are selected (i.e., the ones which are the closest both for the analyzed image and for the ideal template image), and with a dominant orientation (the angle of rotation around the keypoint).

5. Using the selected set of keypoint pairs a projective transformation is estimated using RANSAC. If the projective transformation is estimated and well-defined, the number of inlier pairs is calculated for it (considered zero otherwise).

6. The result is the document template type with the highest number of inlier pairs.

Such an approach allows not only to identify the type of the document, but also to precisely locate the coordinates of the document vertices using the found projective transformation matrix.

Within the scope of the benchmark evaluation, in order to maximize reproducibility, the search for direct and inverse matching was performed without any randomized approximation, and no additional geometric restrictions were imposed, following the procedure described in [13]. For keypoints detection the SURF method was employed

[45], and for matching the keypoints we evaluated the SURF descriptor [45], and the binary descriptor BEBLID

[46] in 256- and 512-dimension variants. For the evaluation, the ground truth quadrangles present in MIDV-2020 annotation for each type of image (except ideal document images) were used, as well as a natural separation of the dataset by the document type, which serves as a ground truth for document identification task. The accuracy of document type identification as measured on MIDV-2020 dataset is presented in tab. 8.

Tab. 8. Feature-based document identification accuracy

Descriptor MIDV-2020 subset

Upright Rotated Photos Video

scans scans frames

SURF 100.00 100.00 95.10 64.38

BEBLID-256 100.00 99.90 98.20 81.75

BEBLID-512 100.00 100.00 98.70 84.48

For an end-to-end system it is important not only to correctly identify the document type, but to correctly find the coordinates of the document's vertices as well, especially if the procedures for text fields extraction and recognition depend on a preliminary rectification of the document page. Tab. 9 presents the benchmark results for cor-

rect type identification and correct location, across the images presented in the MIDV-2020 dataset, using the feature-based approach. Document location quality is estimated using the D4 < 0.02 criterion, introduced in section 2.1.

Tab. 9. Feature-based document identification and location (D4 < 0.02)

Descriptor MIDV-2020 subset

Upright scans Rotated scans Photos Video frames

SURF 98.00 96.70 89.00 54.50

BEBLID-256 99.00 97.70 93.10 72.37

BEBLID-512 98.80 99.00 95.70 75.13

Fig. 7 shows an ideal template image with template keypoints, and in fig. 8 the quadrangle location results of the evaluated methods are presented.

I jiff ' - :■«.'.

Anniz

0mm. w

jt...,

^—---^ Jti£?3 CAfiLOS

, J1:. ! : h

04 05 1V7Ö A « ji 1 9 5 9 4 3 C>'Tl"' 2:024

T r

>tv~

Fig. 7. Ideal document template (from MIDV-500). Yellow points represent the keypoints in variable data zones, which are not used for document location and identification. Green points represent static element keypoints, their neighborhoods are used for descriptors computation

Fig. 8. Analyzed image (photo esp_id/44). Yellow points represent all detected keypoints. Document location results and keypoint inliers are shown in color (SURF — blue, BEBLID-256 - cyan, BEBLID-512 — magenta)

Low accuracy of the feature-based document location and identification on the video frames of MIDV-2020 in comparison with photos and scans could be due to the effects of defocus and motion blur on the ability to dis-

tinguish between local features around small static text elements. The differences between the measured accuracy for different descriptors correlate with the local patch matching accuracy measured in [46].

2.4. Text_ fields recognition

One of the most important tasks related to identity document analysis is the recognition of text fields, which are used for person authentication, identity verification, and automated form filling. In a focused text field recognition scenario, we assume that the preliminary identity document analysis steps, such as detecting and locating the document in an image, and segmenting the document into the separate fields, had already been performed. To provide a baseline for text fields recognition, we employed Tesseract OCR 4.1.1 [47, 48] that provides LSTM-based recognizer for a variety of languages. Previous studies [18] show that Tesseract OCR produces reasonable results in comparison with other popular baseline systems, e.g., Abbyy FineReader [49].

In order to perform an isolated focused text field recognition evaluation, not only the text field values annotations have to be used, but their geometrical annotations as well. To acquire text field images without spatial distortions introduced by the previous document processing steps (document detection, per-field segmentation, etc.), we used the field bounding box coordinates provided in the dataset ground truth for each unique identity document, as well as the document boundary coordinates provided for each photo, scan, and video frame. To obtain rectified field images we extended the annotated bounding boxes by 20 % in each direction. If the field contained letters that cross the baseline or the cap line (this information is provided in the ground truth as additional attributes of each annotated field), we applied an additional 10 % padding.

We divided all fields in the documents into four groups - numerical fields, fields written using numerals and letters of the English alphabet (ISO basic Latin alphabet), fields written using numerals and local alphabets, and machine-readable zones (MRZ). We excluded four fields («authority», «code», «nationality», and «birth_place_eng») of the passport of Greece from our experiment as they contain bilingual (Greek and English) text. The separation of MIDV-2020 into the field groups is specified in tab. 1 of the Appendix.

For each field group, we calculated per-string recognition rate (PSR) and per-character recognition rate (PCR) [18], defined as follows:

PSR=

Lc<

L\o

(5)

where Ltotal denotes the total number of text line images, and Lcorrect stands for the number of correctly recognized text lines;

Motal

^^ min yleV(^ideal ,1irecog ),len(1iideal ) |

PCR = 1 - -¡=-:-,

¿total

S len(/-ideal)

(6)

where len (liideal) denotes the length of the i - th text line, and lev (liideal, lirecog ) is the Levenshtein distance between the recognized text and the ideal text in the annotations.

For the evaluation we did not modify the Tesseract recognition results and annotations in any way, except for the conversion of multiple consecutive spaces into a single space character. The obtained evaluation results are presented in tab. 10.

Tab. 10. Per-field recognition accuracy (Tesseract v4.1.1)

Document code Field groups

Latin alphabet Numerical fields Local alphabets MRZ

PSR PCR PSR PCR PSR PCR PSR PCR

alb id 27.91 62.57 46.73 70.10 20.80 56.07 n/a n/a

aze_passport 37.57 64.85 22.98 46.19 18.78 54.63 4.51 51.34

esp id 28.13 54.97 46.31 68.94 51.72 68.68 n/a n/a

est id 34.64 57.36 35.67 68.64 77.16 87.46 n/a n/a

fin id 50.83 54.97 14.73 55.99 57.36 73.83 n/a n/a

grc_passport 47.52 70.76 38.71 57.86 49.91 68.97 9.64 64.29

lva_passport 35.89 60.25 39.26 67.32 33.84 62.10 12.61 62.69

rus internalpassport n/a n/a 49.54 82.78 43.67 70.74 n/a n/a

srb_passport 36.99 49.80 21.91 44.29 22.38 57.40 15.28 58.46

svk id 68.42 85.89 8.86 51.18 50.35 76.79 n/a n/a

All documents 39.44 63.72 29.19 59.31 37.53 62.49 10.49 59.22

The results in Table 10 correlate with previously obtained results [18] for MIDV-500. The results for MRZ are slightly better than those in [18]. In our view, it is connected with the improvement of Tesseract OCR, since we used version 4.1.1, and the authors of [18] used version 4.0.0. The other field groups cannot be so easily compared as we divide the fields in a different way, e.g., we take all the fields printed with Latin alphabet into the corresponding group, and in [18] only names and last names were included. But we still can say that MIDV-2020 presents more possibilities for text line recognition algorithms improvement, as (i) for 9 out of 10 documents the results in the Latin group are lower than results for the Latin names group in [18]; and (ii) in this paper we presented baseline results for local alphabets, which were not provided before.

2.5. Face detection

In the past two decades, the advancement of technology in electronics and computer science allows the widespread deployment of biometric systems in real-life applications, such as online payment, smartphone-based authentication, biometric passport etc. The face recognition-based biometric systems authorize the legitimate access of the clients by comparing an individual's live face to the face image found in his / her identity docu-

i=i

ment. Consequently, counterfeiting of identity documents has also become a main form to attack face recognition-based biometric systems. Thus, either for the face recognition-based authentication with the identity document or the detection of the counterfeits, the face detection / recognition are indispensable. Moreover, face detection is the first step to start the whole processing.

Instead of detecting live faces under spontaneous scenarios as in [50, 51], detecting face on identity documents is a more challenging problem due to the conditions under which the images of identity documents have been captured being more complex and with much less constraints. The face detection on identity documents encounters more challenges such as partial face, defocus, reflection, the varying shift in space, unconstrained shooting angle and the varying illumination conditions as shown in fig. 9.

Fig. 9. Challenging conditions for capturing the identity documents, as shown in MIDV-2020 dataset: a) partial face (frame svk_id/66/000145); b) defocus (frame alb_id/55/000019); c) reflection (frame fin_id/90/000253); d) shift in space (frame rus_internalpassport/87/000301); e) unconstrained shooting angle (frame rusJnternalpassport/70/000133); andf) bad illumination (frame lva_passport/19/000103)

Thanks to the great advances in deep neural network architectures, face detection in general face images has made great progress in the past decade [52 - 61]. FDDB [50] and WiderFace [51] are the most widely used datasets for training and evaluating face detection algorithms. However, FDDB and WiderFace only include the general / liveness face images in the spontaneous scenarios having no face images from the identity documents. Although FDDB and WiderFace contain the faces captured under the «unconstrained» environment, the face images on the identity documents included in the proposed MIDV-2020 have more variations as shown in fig. 9. Besides, the resolution of images in FDDB and

WiderFace are relatively low (from 40 x 40 to 400 x 400 maximum) compared to the images in MIDV-2020 (2160 x 3840). Thus, not only for the research of face detection on identity documents but also for the research of the generalization of face detection algorithms, the MIDV-2020 is an essential supplement for the current face detection datasets.

Benefiting from the blooming of the deep learning-based generic object detection, face detection makes great progress on the extreme and real variation problem including scale, pose, occlusion, expression, makeup, illumination, blur etc. [55]. Cascade-CNN [56] firstly proposed the deep Cascade-CNNs architecture for face detection. Based on the deep cascaded multi-task framework, MTCNN (MultiTask Cascaded Convolutional Neural Networks) [57] achieved the state-of-the-art performance at that moment, which attained very high speed for face detection and alignment. PCN (Progressive Calibration Network) [58] has tackled the detection of arbitrary rotation-in-plane faces in a coarse-to-fine manner. Based on SSD (Single Shot multibox Detector) [59], S3FD (Single Shot Scale-invariant Face Detector) [60] proposed a scale compensation anchor matching strategy aiming to detect the small faces. RetinaFace [61] employed de-formable context modules and additional landmark annotations to improve the performance of face detection. TinaFace [55] achieved impressive results on unconstrained face detection by revising a generic object detector. However, nowadays the face detection methods seem to make the overall algorithms and systems become more and more complex. In order to provide a strong but simple baseline method for face detection in the images of identity documents, MTCNN is selected in this work as the face detection method. MTCNN is almost the most widely used face detection method of today given its high stability, low latency and simplicity.

The evaluation of face detection on MIDV-2020 has been conducted on the five different types of images of identity documents respectively. From the perspective of identity document analysis, the personal photo as a document field might be considered as a separate entity from the face, mainly due to different goals of its extraction from the document and of its further analysis. Thus, in the annotations provided with the MIDV-2020 dataset for each unique document the bounding box of a personal photo area is provided for each template image, as well as the bounding box of a face for each template image, photo, scan, and video frame. The ground truth of bounding box of faces in MIDV-2020 has been semi-automatically annotated, i.e., the faces in the dataset have been firstly detected automatically by pretrained MTCNN and then the detected bounding boxes have been calibrated manually to serve as the ground truth of the bounding boxes of faces in MIDV-2020. In particular, there are more than one face in some identity documents: the principal face with a large scale and the other watermark-like tiny faces as shown in fig. 1. In this work, we only adopted the

principal face to employ the evaluation. The MTCNN model used in this work has been pretrained on WiderFace. Thus, we can evaluate the generalization capacity of a face detection model trained on the general images performing on the heterogeneous document images. The evaluation has been conducted on one GPU (Nvidia GeForce RTX 2080).

As being an application of generic object detection, the face detection evaluation also follows the evaluation protocol of generic object detection. The widely used detection evaluation metrics Average Precision (AP) [62] is adopted to evaluate the performance of MTCNN on document images. Instead of having several different classes of objects to detect, face detection has only one class to detect. Thus, the AP here is the average precision of the detection of different faces. As same as in [62], three types of AP with different Intersection over Union (IoU) values have been used as three types of metrics for evaluating the face detection method. The metric with three different IoU thresholds indicates three different levels to locate the faces in images. The IoU threshold from 0.3 to 0.7 indicates the localization requirement from easy to hard. That means AP with IoU of 0.7 reward detectors with better localization than AP with IoU of 0.5. The source code for the evaluation scripts are available at [63].

Tab. 11 shows the evaluation results of face detection on MIDV-2020 by MTCNN. Generally speaking, MTCNN has obtained a good detection result in every type of identity document images. Only for the scan copies of rotated identity cards, the performance of MTCNN is inferior to the others especially when IoU threshold is 0.7 since MTCNN is not designed for detecting the rotated face as PCN. Nevertheless, MTCNN still shows a comparable performance for detecting rotated faces when the localization requirement is less strict such as the AP with IoU threshold of 0.5 or 0.3. We can see that even MTCNN is pretrained on the general face images in WiderFace, it can still well generalize on the heterogeneous document images.

Tab. 11. The face detection performance of MTCNN on MIDV-2020

MIDV-2020 Average Precision (AP)

subset Easy (IoU > 0.3) Medium (IoU > 0.5) Hard (IoU > 0.7)

Templates 0.99899 0.99899 0.99899

Upright 1.00000 1.00000 1.00000

scans

Rotated 0.99057 0.99057 0.90178

scans

Photos 0.99403 0.99303 0.98865

Video frames 0.98731 0.98657 0.96160

Fig. 10 and Fig. 11 shows the ROC curves and Precision-Recall curve of face detection on different types of documents under different localization requirements, i.e., with different IoU thresholds.

From the fig. 10 and fig. 11, we can see that the performance of MTCNN on scan copies of identity docu-

ments in the upright orientation («Upright scans» subset, green curve) and templates («Templates» subset, violet curve) are superior to the others, since the identity documents in these two types are always placed properly without any shift, reflect, occlusion etc. and the images have been captured in the perfect condition. There is almost no false detection in these two types. Instead, the face detection on scan copies of rotated documents («Ro-tated scans» subset, orange curve) shows relatively worse result, which is more pronounced in fig. 11.

True Positive Rate

1.0

0.8 0.6 0.4 0.2

—i-1-1-1-1-

i i i i i ----

i i

i i i i i i i i i i

i i i i i i i i

i -Video frames Rotated scan -Upright scan -Photo -Templates

(a)

0 200 True Positive Rate

400

600 800

False Positive

0.8 0.6 0.4 0.2

-------- ------- --------

r n -Video frames Rotated scan -Upright scan -Photo -Templates

------- -------- ------

(b)

0

200

400

600 800

False Positive

True Positive Rate 1.0

0.8

0.6

0.4

0.2

i i i i i i

i i i i i i ____

1 1 1 1 1 1 1 1 1 1

1 1 1 1 1 1 1 1

■ - -Video frames Rotated scan -Upright scan -Photo — Templates

(c)

0

500

1000 1500

2000 2500 False Positive

Fig. 10. ROC curves of MTCNN detecting faces of MIDV-2020 with three configurations: (a) easy mode with IoU threshold 0.3; (b) medium mode with IoU threshold 0.5; (c) hard mode with IoU threshold 0.7

Precision 1.000

0.998

0.996

0.994

0.992

0.990

0.988

0.986

—__ \ \ —, * ,-- v—— \ T 1 1 1 I 1 I 1 1

_____ ______ 1 1 1 ✓il

r r 1 1 rr i ] I

1 1 1

----- --------p 1 1 ________1__ -Video frames Rotated scan -Upright scan -Photo -Templates I j_____

1 1 1 1 1

(a)

0.2

0.4

0.6

0.8

1.0 Recall

Precision

1.0 Recall

Precision

0.980 0.960 0.940 0.920 0.900

,

1

— Video frames Rotated scan -Upright scan -Photo -Templates

I *

(c)

0.2

0.4

0.6

0.8

1.0

Recall

Fig. 11. Precision-Recall curves of MTCNN detecting faces of MIDV-2020 with three configurations: (a) easy mode with IoU threshold 0.3; (b) medium mode with IoU threshold 0.5; (c) hard mode with IoU threshold 0.7

We can see that MTCNN performs well for detecting faces on the document images as a baseline model pre-trained on general images. Even for some hard cases, MTCNN can still detect the faces on the documents (see Figure 12). Moreover, the performance of MTCNN on MIDV-2020 is even better than MTCNN detecting the faces on general images of WiderFace (i.e. 0.85 on easy set) which is used to train MTCNN. It shows the stability of MTCNN even on the heterogeneous document images. It has to point out that there are two reasons may help MTCNN to obtain such good results. First, the faces presented on the documents are frontal faces no matter how

the documents have been placed in space. Comparing to the profile or the faces with extreme pose included in WiderFace, the frontal faces are much easy to detect. Secondly, the ground truth of bounding boxes of faces in MIDV-2020 were semi-automatically annotated with MTCNN, which may also alleviate the detection error of MTCNN. Nevertheless, for the extreme complex cases such as the images with the strong reflect, extreme shooting angle or partial face as shown in Figure 12, the general face detection model still fails to detect the faces. These extreme hard cases have seldom appeared in the general face detection datasets such as FDDB or WiderFace. Thus, MIDV-2020 provides a useful supplementary material to the current face analysis benchmarks enabling the face detection algorithms to be generalized in more different scenarios or domains.

Fig. 12. Hard cases for the face detection in MIDV-2020; hard cases which were successfully detected using MTCNN: (a) high

rotation angle (frame alb_id/34/000001) and (b) bad illumination (frame esp_id/19/000001); extremely hard cases which MTCNN failed to detect: (c) strong reflection (frame srb_passport/91/000181), (d) extreme shooting angle (frame svk_id/71/000007), (e) partial face (frame svk_id/66/000145)

Conclusion

In this article, we presented a new identity documents dataset MIDV-2020, which consists of 2000 scans, 1000 photos, and 1000 video clips of different artificially generated identity documents. In total, the dataset contains 72409 annotated images, which makes it the largest publicly available dataset dedicated to identity documents. We have presented baseline evaluations of the methods of such identity document image analysis problems as doc-

ument detection, location, and identification; text fields recognition; and face detection. The first group of document location methods - the ones which do not rely on the document page content, instead only analyse the boundaries, perform well on upright scans, but even rotational changes in the document position significantly degrade the location precision. Challenges related to mobile capture, such as projective distortions, degrade the accuracy and performance further, thus the need for improvements are clear when it comes to ID document analysis use cases with variable input sources. The evaluated semantic segmentation-based methods do not exhibit improvement over boundary-based methods, presumably due to high dependence on the training data and high sensitivity to background, however this should be investigated further. Feature-based methods of document location and identification perform well on scanned images and photos, however still leave a lot of room for improvements when evaluated on video frames. The text field recognition results obtained in this article correspond to the ones previously observed for the dataset MIDV-500 from the same domain, however the evaluation on MIDV-2020 presents baselines for local language groups, which were not previously evaluated for MIDV-500. While the task of text line recognition is a well studied one, the obtained results show that given modern challenges the text recognition methods should be significantly improved for practical and robust ID document recognition systems. Modern deep learning-based methods of face detection were shown to perform with high precision in different conditions, however there exist complex cases related to sharp camera angles and, partial face visibility, and highlights, which still present challenges. Overall, the performed experiments show that the dataset contains examples of images and video clips captured in conditions which present challenges for existing analysis methods, and which can serve as a motivation for further improvement of document analysis methods and algorithms.

Although the MIDV-2020 dataset is in itself a dataset of fake documents, its usage for developing and evaluating methods of computational document forensics, presentation attacks detection, and other authenticity validation approaches, is limited. As a future work, we plan to expand the dataset to include capturing conditions and document features, which would allow training and evaluation of fraud prevention techniques, such as the analysis of document color characteristics, holographic security layer, and optically variable devices [64, 65].

Authors believe that the dataset will serve as a benchmark for document analysis methods and identity document processing systems, and prove valuable for the advancement of the field of document analysis and recognition in general. The dataset is publicly available for download at [66] and [67], the alternative mirrors have identical content.

References

[1] Fang X, Fu X, Xu X. ID card identification system based on image recognition. 12th IEEE Conf on Industrial Electronics and Applications (ICIEA) 2017: 1488-1492. DOI: 10.1109/ICIEA.2017.8283074.

[2] Attivissimo F, Giaquinto N, Scarpetta M, Spadavecchia M. An automatic reader of identity documents. IEEE Int Conf on Systems, Man and Cybernetics (SMC) 2019: 35253530. DOI: 10.1109/SMC.2019.8914438.

[3] Kuklinski T, Monk B. The use of ID reader-authenticators in secure access control and credentialing. IEEE Conf on Technologies for Homeland Security 2008: 246-251. DOI: 10.1109/THS.2008.4534458.

[4] Soares A, das Neves Junior R, Bezerra B. BID Dataset: a challenge dataset for document processing tasks. Anais Estendidos do XXXIII Conf on Graphics, Patterns and Images 2020: 143-146. D0I:10.5753/sibgrapi.est.2020.12997.

[5] Ghanmi N, Nabli C, Awal AM. CheckSim: A reference-based identity document verification by image similarity measure. In Book: Smith EHB, Pal U, eds. Document analysis and recognition - ICDAR 2021 Workshops. Springer Nature Switzerland AG; 2021: 422-436. DOI: 10.1007/978-3-030-86198-8_30.

[6] Chiron G, Arrestier F, Awal AM. Fast end-to-end deep learning identity document detection, classification and cropping. In Book: Lladôs J, Lopresti D, Uchida S, eds. Document analysis and recognition - ICDAR 2021. Springer Nature Switzerland AG; 2021: 333-347. DOI: 10.1007/978-3-030-86337-1_23.

[7] Ngoc MOV, Fabrizio J, Geraud T. Saliency-based detection of identity documents captured by smartphones. 12th IAPR Int Workshop on Document Analysis Systems (DAS) 2018: 387-392. DOI: 10.1109/DAS.2018.17.

[8] Arlazarov VV, Bulatov K, Chernov T, Arlazarov VL. MIDV-500: a dataset for identity document analysis and recognition on mobile devices in video stream. Computer Optics 2019; 43(5): 818-824. DOI: 10.18287/2412-61792019-43-5-818-824.

[9] Bulatov K, Matalov D, Arlazarov VV. MIDV-2019: challenges of the modern mobile-based document OCR. Proc SPIE 2020; 11433: 114332N. DOI: 10.1117/12.2558438.

[10] Chazalon J, Gomez-Kramer P, Burie J-C, Coustaty M, Eskenazi S, Luqman M, Nayef N, Rusinol M, Sidere N, Ogier J. SmartDoc 2017 Video Capture: mobile document acquisition in video mode. 14th IAPR Int Conf on Document Analysis and Recognition (ICDAR) 2017; 4: 11-16. DOI: 10.1109/ICDAR.2017.306.

[11] Skoryukina N, Arlazarov VV, Nikolaev D. Fast method of ID documents location and type identification for mobile and server application. Int Conf on Document Analysis and Recognition (ICDAR) 2019: 850-857. DOI: 10.1109/ICDAR.2019.00141.

[12] Buonanno A, Nogarotto A, Cacace G, di Gennaro G, Palmieri FAN, Valenti M, Graditi G. Bayesian feature fusion using factor graph in reduced normal form. Appl 2021; 11(4): 1934. DOI: 10.3390/app11041934.

[13] Skoryukina N, Faradjev I, Bulatov K, Arlazarov VV. Impact of geometrical restrictions in RANSAC sampling on the ID document classification. Proc SPIE 2020; 11433: 1143306. DOI: 10.1117/12.2559306.

[14] Lynchenko A, Sheshkus A, Arlazarov VL. Document image recognition algorithm based on similarity metric robust to projective distortions for mobile devices. Proc SPIE 2019; 11041: 110411K. DOI: 10.1117/12.2523152.

[15] das Neves Junior RB, Lima E, Bezerra BL, Zanchettin C, Toselli AH. HU-PageScan: a fully convolutional neural network for document page crop. IET Image Proces 2020; 14: 3890-3898. DOI: 10.1049/iet-ipr.2020.0532.

[16] Sheshkus A, Nikolaev D, Arlazarov VL. Houghencoder: neural network architecture for document image semantic segmentation. IEEE Int Conf on Image Processing (ICIP), 2020: 1946-1950. DOI: 10.1109/ICIP40778.2020.9191182.

[17] Bakkali S, Luqman MM, Ming Z, Burie J. Face detection in camera captured images of identity documents under challenging conditions. Int Conf on Document Analysis and Recognition Workshops (ICDARW) 2019: 55-60. DOI: 10.1109/ICDARW.2019.30065.

[18] Chernyshova YS, Sheshkus AV, Arlazarov VV. Two-step CNN framework for text line recognition in camera-captured images. IEEE Access 2020; 8: 32587-32600. DOI: 10.1109/ACCESS.2020.2974051.

[19] Petrova O, Bulatov K, Arlazarov VV, Arlazarov VL. Weighted combination of per-frame recognition results for text recognition in a video stream. Computer Optics 2021; 45(1): 77-89. DOI: 10.18287/2412-6179-CO-795.

[20] Bulatov KB. A method to reduce errors of string recognition based on combination of several recognition results with per-character alternatives [In Russian]. Vestnik Yu-zhno-Ural'skogo Universiteta. Seriya Matematicheskoe Modelirovanie i Programmirovanie 2019; 12(3): 74-88. DOI: 10.14529/mmp190307.

[21] Bulatov K, Razumnyi N, Arlazarov VV. On optimal stopping strategies for text recognition in a video stream as an application of a monotone sequential decision model. Int J Doc Anal Recognit 2019; 22(3): 303-314. DOI: 10.1007/s10032-019-00333-0.

[22] Bulatov K, Fedotova N, Arlazarov VV. Fast approximate modelling of the next combination result for stopping the text recognition in a video. 25th Int Conf on Pattern Recognition (ICPR) 2021: 239-246. DOI: 10.1109/ICPR48806.2021.9412574.

[23] Aliev MA, Kunina IA, Kazbekov AV, Arlazarov VL. Algorithm for choosing the best frame in a video stream in the task of identity document recognition. Computer Optics 2021; 45(1): 101-109. DOI: 10.18287/2412-6179-CO-811.

[24] Chernov TS, Ilyuhin SA, Arlazarov VV. Application of dynamic saliency maps to video stream recognition systems with image quality assessment. Proc SPIE 2019; 11041: 110410T. DOI: 10.1117/12.2522768.

[25] Myasnikov E, Savchenko A. Detection of sensitive textual information in user photo albums on mobile devices. Int Multi-Conf on Engineering, Computer and Information Sciences (SIBIRCON) 2019: 384-390. DOI: 10.1109/SIBIRCON48586.2019.8958325.

[26] Castelblanco A, Solano J, Lopez C, Rivera E, Tengana L, Ochoa M. Machine learning techniques for identity document verification in uncontrolled environments: a case study. In Book: Mora KMF, Marín JA, Cerda J, Carrasco-Ochoa JA, José Francisco Martínez-Trinidad, José Arturo Olvera-López, eds. Pattern recognition. Proceedings. Springer Nature Switzerland AG; 2020: 271-281. DOI: 10.1007/978-3-030-49076-8_26.

[27] Council of the European Union. PRADO - Public Register of Authentic identity and travel Documents Online, Source: (https://www.consilium.europa.eu/prado).

[28] Wikipedia. Category: Serbian masculine given names. Source:

(https://en.wikipedia.org/wiki/Category:Serbian_masculine _given_names).

[29] Fantasy name generators: Azerbaijani names. Source: (https://www.fantasynamegenerators.com/azerbaijani-names.php).

[30] Generated Photos. Source: (https://generated.photos).

[31] Karras T, Laine S, Aila T. A style-based generator architecture for generative adversarial networks. IEEE/CVF Conf on Computer Vision and Pattern Recognition (CVPR) 2019: 4396-4405. DOI: 10.1109/CVPR.2019.00453.

[32] Dutta A, Zisserman A. The VIA annotation software for images, audio and video. Proc 27th ACM Int Conf on Multimedia (MM'19) 2019: 2276-2279. DOI: 10.1145/3343031.3350535.

[33] VGG Image Annotator (VIA). Source: (https://www.robots.ox.ac.uk/~vgg/software/via).

[34] Javed K, Shafait F. Real-time document localization in natural images by recursive application of a CNN. 14th IAPR Int Conf on Document Analysis and Recognition (ICDAR) 2017: 105-110. DOI: 10.1109/ICDAR.2017.26.

[35] Zhu A, Zhang C, Zhi L, Xiong S. Coarse-to-fine document localization in natural scene images with regional attention and recursive corner refinement. Int J Doc Anal Recognit 2019; 22: 351-360. DOI: 10.1007/s10032-019-00341-0.

[36] Skoryukina N, Nikolaev DP, Sheshkus A, Polevoy D. Real time rectangular document detection on mobile devices. Proc SPIE 2015; 9445: 94452A. DOI: 10.1117/12.2181377.

[37] Tropin DV, Ershov AM, Nikolaev DP, Arlazarov VV. Advanced Hough-based method for on-device document localization. Computer Optics 2021; 45(5): 702-712. DOI: 10.18287/2412-6179-CO-895.

[38] Tropin DV, Ilyuhin SA, Nikolaev DP, Arlazarov VV. Approach for document detection by contrours and contrasts. 25th Int Conf on Pattern Recognition (ICPR) 2021: 96899695. DOI: 10.1109/ICPR48806.2021.9413271.

[39] Tropin DV, Konovalenko IA, Skoryukina NS, Nikolaev DP, Arlazarov VV. Improved algorithm of ID card detection by a priori knowledge of the document aspect ratio. Proc SPIE 2021; 11605: 116051F. DOI: 10.1117/12.2587029.

[40] Ngoc MOV, Fabrizio J, Geraud T. Document detection in videos captured by smartphones using a saliency-based method. Int Conf on Document Analysis and Recognition Workshops (ICDARW) 2019: 19-24. DOI: 10.1109/ICDARW.2019.30059.

[41] Burie J, et al. ICDAR2015 competition on smartphone document capture and OCR (SmartDoc). 13th Int Conf on Document Analysis and Recognition (ICDAR) 2015: 1161-1165. DOI: 10.1109/ICDAR.2015.7333943.

[42] Liu L, Wang Z, Qiu T, Chen Q, Lu Y, Suen CY. Document image classification: Progress over two decades. Neurocomputing 2021; 453: 223-240. DOI: 10.1016/j.neucom.2021.04.114.

[43] Chiron G, Ghanmi N, Awal AM. ID documents matching and localization with multi-hypothesis constraints. 25th Int Conf on Pattern Recognition (ICPR) 2021: 3644-3651. DOI: 10.1109/ICPR48806.2021.9412437.

[44] Awal AM, Ghanmi N, Sicre R, Furon T. Complex document classification and localization application on identity document images. 14th IAPR Int Conf on Document Anal-y sis and Recognition (ICDAR) 2017: 426-431. DOI: 10.1109/ICDAR.2017.77.

[45] Bay H, Ess A, Tuytelaars T, Gool LV. Speeded-up robust features (SURF). Comput Vis Image Underst 2008; 110(3): 346-359. DOI: 10.1016/j.cviu.2007.09.014.

[46] Suarez I, Sfeir G, Buenaposada JM, Baumela L. BEBLID: boosted efficient binary local image descriptor. Patt Recogn Lett 2020; 133: 366-372. DOI: 10.1016/j.patrec.2020.04.005.

[47] Smith R. An overview of the Tesseract OCR engine. 9th Int Conf on Document Analysis and Recognition (ICDAR) 2007: 629-633. DOI: 10.1109/ICDAR.2007.4376991.

[48] Smith R, Podobny Z, et al. Tesseract OCR. Source: (https://github.com/tesseract-ocr/tesseract).

[49] ABBYY FineReader PDF: the smarter PDF solution. Source: (https: //pdf. abbyy.com).

[50] Jain V, Learned-Miller E. FDDB: A benchmark for face detection in unconstrained settings. Amherst: University of Massachusetts; 2010: UM-CS-2010-009. Source: (http://vis-www.cs.umass.edu/fddb).

[51] Yang S, Luo P, Loy CC, Tang X. WIDER FACE: a face detection benchmark. IEEE Conf on Computer Vision and Pattern Recognition (CVPR) 2016: 5525-5533. DOI: 10.1109/CVPR.2016.596.

[52] Hu P, Ramanan D. Finding tiny faces. IEEE Conf on Computer Vision and Pattern Recognition (CVPR) 2017: 1522-1530. DOI: 10.1109/CVPR.2017.166.

[53] Najibi M, Samangouei P, Chellappa R, Davis LS. SSH: single stage headless face detector. IEEE Int Conf on Computer Vision (ICCV) 2017: 4885-4894. DOI: 10.1109/ICCV.2017.522.

[54] Yang S, Luo P, Loy C, Tang X. From facial parts responses to face detection: a deep learning approach. IEEE Int Conf on Computer Vision (ICCV) 2015: 3676-3684. DOI: 10.1109/ICCV.2015.419.

[55] Zhu Y, Cai H, Zhang S, Wang C, Xiong Y. TinaFace: strong but simple baseline for face detection. arXiv Preprint. Source: (https://arxiv.org/abs/2011.13183).

[56] Li H, Lin Z, Shen X, Brandt J, Hua G. A convolutional neural network cascade for face detection. IEEE Conf on Computer Vision and Pattern Recognition (CVPR) 2015: 5325-5334. DOI: 10.1109/CVPR.2015.7299170.

[57] Zhang K, Zhang Z, Li Z, Qiao Y. Joint face detection and alignment using multitask cascaded convolutional networks. IEEE Signal Proces Lett 2016; 23(10): 1499-1503. DOI: 10.1109/LSP.2016.2603342.

[58] Shi X, Shan S, Kan M, Wu S, Chen X. Real-time rotationinvariant face detection with progressive calibration networks. IEEE Conf on Computer Vision and Pattern Recognition (CVPR) 2018: 2295-2303. DOI: 10.1109/CVPR.2018.00244.

[59] Liu W, Anguelov D, Erhan D, Szegedy C, Reed S, Fu C-Y, Berg AC. SSD: single shot multibox detector. In Book: Leibe B, Matas J, Sebe N, Welling M, eds. Computer Vision - ECCV 2016. Cham: Springer International Publishing AG; 2016: 21-37. DOI: 10.1007/978-3-319-46448-0_2.

[60] Zhang S, Zhu X, Lei Z, Shi H, Wang X, Li SZ. S3FD: single shot scale-invariant face detector. IEEE Int Conf on Computer Vision (ICCV) 2017: 192-201. DOI: 10.1109/ICCV.2017.30.

[61] Deng Z, Guo J, Ververas E, Kotsia I, Zafeiriou S. Ret-inaFace: single-stage dense face localisation in the wild. IEEE Conf on Computer Vision and Pattern Recognition (CVPR) 2020: 5203-5212. DOI: 10.1109/CVPR42600.2020.00525.

[62] Lin TY, Maire M, Belongie S, Hays J, Perona P, Ramanan D, Dollar P, Zitnick CL. Microsoft COCO: common objects in context. In Book: Fleet D, Pajdla T, Schiele B, Tuytelaars T, eds. Computer Vision - ECCV 2014. Switzerland: Springer International Publishing; 2014. DOI: 10.1007/978-3-319-10602-1_48.

[63] The source code for average precision metric. Source: (https://github.com/hengxyz/MIDV_2020_det).

[64] Hartl AD, Arth C, Grubert, J, Schmalstieg D. Efficient verification of holograms using mobile augmented reality. IEEE Trans Vis Comput Graph 2015; 22(7): 1843-1851. DOI: 10.1109/TVCG.2015.2498612.

[65] Polevoy DV, Panfilova EI, Nikolaev DP. White balance correction for detection of holograms in color images of black and white photographs [In Russian]. Journal of Information Technologies and Computing Systems 2021; 3: 82-95. DOI: 10.14357/20718632210308.

[66] MIDV-2020 dataset. Mirror 1. Source: (ftp://smartengines.com/midv-2020).

[67] MIDV-2020 dataset. Mirror 2. Source: (http://l3i-share.univ-lr.fr).

Appendix

Tab. 1. Specification of MIDV-2020fields according to the field groups

Document type Latin alphabet Numerical fields Local alphabets MRZ

alb_id authority; gender; id number; nationality birth date; issue date; expiry date; numbers birth_place; name; surname mrz line0; mrz linel

aze_passport type; code; number; surname eng; name eng; id number; gender; authority eng birth date; issue date; expiry date; expiry date 2 surname; name; nationality/nationality eng; birth_place; authority

esp id id number; gender; nationality; number; name code birth date; expiry date; issue date name; surname; surname second

est id name; gender; nationality; number; number2; number3 birth date; id number; expiry date surname

fin id gender; nationality number; birth date; issue date; expiry date; birth date 2; birth date 22 surname; name; code

grc_passport birth date; birth country; expiry date; gender; issue date; name eng; surname eng; type; number height birth_place; name; surname

lva_passport code; gender; nationality; number; type birth date; expiry date; height; id number; issue date authority line0; authority line1; birth place; name; surname

rus internalpass port birth date; number birth_place line0; birth_place line1; birth_place line2; gender; name; patronymic; surname

srb_passport code; nationality; type birth date; expiry date; id number; issue date; number; expiry date2 authority; birth_place; gender; name; residence line0; residence line1; surname

svk id gender; nationality; number id number; birth date; expiry date; issue date name; surname; issue_place

/images/ # The directory with all images

/<CODE>/ # Subdirectory name corresponds to type code, see Table 1

00.jpg

01.jpg

99.jpg

/annotations/ # The directory with all image annotations

<CODE>.json # Annotations for each type code is in a single file

Fig. 1. The content structure of the «templates.tar», «scan upright.tar», «scan rotated.tar», «photo.tar» archives /images/ # The directory with original TIFF scans

/<CODE>/ # Subdirectory name corresponds to type code, see Table 1

00.tif 01.tif

99.tif

/images/

/<CODE>/ /00/

Fig. 2. The content structure of the «scan upright tif.tar», «scan rotated tif.tar»

000001.jpg 000007.jpg

# The directory with all annotated clip frames

# Document type codes, see Table 1

# Subdirectory with clip number (00-99)

# Each 6-th frame of each clip is stored

/99/

000001.jpg 000007.jpg

/annotations/ /<CODE>/

00.json

# The directory with clip annotations

# Document type codes, see Table 1

# Annotations for all frames of a single clip

9 9.j s on

Fig. 3. The content structure of the «clips.tar»

/video/ # The directory with all original video clips

/<CODE>/ # Document type codes, see Table 1

00. mp4[MOV] # Video clips have either "mp4" or "MOV" extensions.

99. mp4[MOV]

Fig. 4. The content structure of the «clips video.tar»

"_via_img_metadata": {

"00.jpg1859173": { # Key: image name and size (in bytes)

"filename": "00.jpg", # Image name

"size": 1859173, # Image size (in bytes)

"regions": [ # List of regions

{

"shape_attributes": { # Coordinates of the bounding

# box of a field

"name": "rect",

"x": 708, "y": 727, "width": 360, "height": 52

},

"region_attributes": {

_ # Field annotation

'field_name": "birth_date", # Field name

"value": "20-11-1991" "features": {}

# Field value

# Additional information

"file attributes": {}

# VIA-related metadata

Fig. 5. Example of the document template annotation in JSONformat (base format - VIA v2.0.11 [32, 33])

{

"_via_img_metadata": {

"00.jpg955580": { # Key: image name and size (in bytes)

"filename": "00.jpg", # Image name "size": 955580, # Image size (in bytes)

"regions": [ {

"shape_attributes": { # Coordinates of the face bounding box "name": "rect",

"x": 1296, "y": 1685, "width": 193, "height": 177

},

"region_attributes": {

"field_name": "face", "value": "", "features": {}

}

},

{

"shape_attributes": { # Coordinates of document boundary quadrangle "name": "polygon", # Points are given from the top-left corner of

# the document in clockwise order "all_points_x": [1223, 882, 1480, 1819], # X-coordinates "all_points_y": [2038, 1085, 869, 1820] # Y-coordinates

},

"region_attributes": { "field_name": "doc_quad" }

}

],

"file_attributes": {}

},

},

J_

Fig. 6. Example of the annotation in JSON format for scans, photos, and video frames (base format - VIA v2.0.11 [32, 33])

Authors' information

Konstantin Bulatovich Bulatov, graduated from National University of Science and Technology «MISiS» in 2013, majoring in Applied Mathematics, and received his Ph.D. degree in 2020. Currently he works as a senior researcher a the Federal Research Center «Computer Science and Control» of RAS. Research interests include pattern recognition, combinatorial optimization, computer vision, document analysis. E-mail: [email protected] .

Ekaterina Vladimirovna Emelianova, graduated from National University of Science and Technology «MISiS» in 2021, majoring in Applied Mathematics, and Lighting and Industrial Design. Research interests include computer graphics processing and graphics design. E-mail: [email protected] .

Daniil Vyacheslavovich Tropin, graduated from Moscow Institute of Physics and Technology (State University) in 2019, majoring in Applied Mathematics and Informatics. Currently he is a postgraduate student at the latter and lead programmer at the Federal Research Center «Computer Science and Control» of RAS. Research interests are image processing and computer vision. E-mail: [email protected] .

Natalya Sergeevna Skoryukina, graduated from National University of Science and Technology «MISiS» in 2013, majoring in Applied Mathematics. Computer programmer (1st category) at the Federal Research Center "Computer Science and Control" of Russian Academy of Sciences. Scientific interests: image analysis, computer vision. E-mail: [email protected] .

Yulia Sergeevna Chernyshova, received the B.S. degree in Applied Mathematics and the M.S. degree in Applied Computer Science from National University of Science and Technology «MISiS», Moscow, Russia, in 2016 and 2018, respectively. She is currently pursuing a Ph.D. degree in computer science at the Federal Research Center «Computer Science and Control» of RAS. Her research interests include training data synthesis, optical character recognition and deep neural networks. E-mail: [email protected] .

Alexander Vladimirovich Sheshkus, received his B.S. and M.S. degrees in Applied Physics and Mathematics from the Moscow Institute of Physics and Technology in 2009 and 2011, respectively. Currently he works as a researcher

with the Federal Research Center «Computer Science and Control» of Russian Academy of Sciences. His research interests include deep neural networks, computer vision, and projective invariant image segmentation. E-mail: [email protected] .

Sergey Alexandrovich Usilin, graduated from Moscow Institute of Physics and Technology in 2009. Received his Ph.D. degree in 2018. Works as a Senior Researcher at the Federal Research Center «Computer Science and Control» of Russian Academy of Sciences. Scope of scientific interests: object detection, machine learning, recognition systems, digital image processing. E-mail: [email protected] .

Zuheng Ming, received Ph.D. in Computer Science from University of Grenoble (France). Zuheng Ming is currently working as post-doc at L3i Laboratory of La Rochelle University (France). His research interests include computer vision, multimodal deep learning and biometric face recognition. E-mail: [email protected] .

Jean-Christophe Burie, received his Ph.D. degree in Automatic Control Engineering and Industrial Data Processing from University of Lille (France) in 1995. He is a Full Professor in Computer Science at L3i Laboratory of La Rochelle University (France). He is currently working on image processing, image and video analysis. His research activities related to IA and deep learning focuses on document analysis and face recognition. E-mail: _ jcburie@,univ-lr.fr .

Muhammad Muzzamil Luqman, received Ph.D. in Computer Science from François Rabelais University of Tours (France) and Autonoma University of Barcelona (Spain). Luqman is currently working on the post of Research Engineer at L3i Laboratory of La Rochelle University (France). His research interests include document image analysis, pattern recognition and computer vision. E-mail: muhammad_muzzamil.luqman@,univ-lr.fr .

Vladimir Viktorovich Arlazarov, graduated from Moscow Institute of Steel and Alloys in 1999, majoring in Applied Mathematics. Received his Ph.D. degree in 2005. Currently he works as a head of division 93 at the Federal Research Center «Computer Science and Control» of RAS. Research interests are pattern recognition, computer vision, intelligent systems, and machine learning. E-mail: [email protected] .

Code of State Categories Scientific and Technical Information (in Russian - GRNTI)): 28.23.15 Received July 1, 2021. The final version - November 18, 2021.