• 专利附录
  • 专利说明
  • 权利要求
  • 类似技术
  • 同族专利
  • 引用文献
  • 法律状态
  • 优先权
    • 专利摘要:
    A method of recognizing characters in an image of a document comprising at least one alphanumeric field, the method comprising the steps of: segmenting the image to identify objects therein; define an enclosing box around each object; performing multiple successive selections based on different descriptors and dividing the bounding boxes into a plurality of cells each of which is determined a texture descriptor as an oriented gradient histogram; perform character recognition on the bounding boxes that are finally selected. Device for implementing this method
    公开号:FR3081245A1
    申请号:FR1854139
    申请日:2018-05-17
    公开日:2019-11-22
    发明作者:Segbedji Goubalan；Thierry Viguier
    申请人:Idemia Identity and Security France SAS；
    IPC主号:
    专利说明:

    The present invention relates to the field of image processing for the purpose of achieving character recognition in any written document such as, for example, a transport ticket or an identity document.
    Background of the invention
    An identity document, such as a passport or a national identity card, contains text fields containing in the form of sequences of alphanumeric characters, for example, name, first names, date and place of birth. of the identity document holder, as well as the name of the authority which issued the identity document and the date of issue.
    Certain administrative operations require having a facsimile of the document and re-entering the content of at least some of these fields. To speed up processing, it is known to scan the document and extract the content from the text fields using a computer program implementing a character recognition algorithm.
    It is also known to add to these ·· documents security elements intended to complicate the falsification and unauthorized reproduction of this document. These security elements · are often present in the background of the document and include for example decorations or fine lines forming · patterns or characters.
    0r, it happens that · these security elements, in particular when: they are strongly contrasted and border a text field, are interpreted as 30 being characters by the character recognition program. This results in errors prejudicial to the efficiency of the image processing applied to the documents and consequently to the completion of the administrative formalities.
    Subject of the invention
    An object of the invention is to provide a means for making the recognition of characters more reliable: in particular when the background is heterogeneous and / or when the background 5 is not known a priori.
    Brief description of the invention
    To this end, provision is made, according to the invention, for a method of recognizing characters in an image of a document comprising at least one alphanumeric field, the method comprising the steps of:
    - segment the image to identify objects in it;
    - define a bounding box around each object and perform a first: selection to select the bounding boxes supposedly containing a character: as a function of at least one theoretical dimensional characteristic of an alphanumeric character;
    make a second selection: comprising the application to each selected bounding box of 20 shape descriptors and the implementation of a decision-making algorithm to select, on the basis of the descriptors, the bounding boxes supposedly containing a character;
    group the bounding boxes according to 25 relative positions of the bounding boxes;
    - make a third selection by dividing each of these bounding boxes into a plurality; cells for each · of which a texture descriptor is determined in the form of an oriented gradient histogram, the histograms then being concatenated and a decision-making algorithm being implemented to select, on the basis of the descriptors, the bounding boxes supposedly containing a character;
    - perform character recognition on the finally selected bounding boxes.
    The process of the invention makes it possible, without human intervention, to limit the influence of the background of the image and of the digitization artifacts on the extraction of the characters: alphanumeric present in the image, which makes it possible to improve the reliability of automatic character recognition. This further allows character recognition even from digitization having a quality which would have been considered insufficient to achieve character recognition using prior art methods.
    The invention also relates to a character recognition device comprising a: computer unit provided with the means for its connection to a digitizing apparatus arranged to perform a digitization of a written document. The computer unit comprises at least one processor and a memory containing a program implementing the method according to the invention.
    Other characteristics and advantages of the invention will emerge on reading the following description of a particular non-limiting embodiment of the invention.
    Brief description of the drawings
    Reference will be made to the attached drawings, among which:
    - Figure 1 is a schematic view of a device for · implementing the method of the invention;
    FIG. 2 is a schematic view of an image of a document comprising characters recognizable by the method according to the invention;
    Figure 3 is a diagram showing the different
    steps of the method according to the invention; - the figures 4. a and 4. b are views of: detail of this image before and after strengthening of contrast;the figures/ 5. a and 5: .b are views of detail of this image before and after / reinforcement of contrast;the figures 5. c and 5.d are views of /detail of this image while the segmentation of 1 ' picture/ at
    using a mask.
    Detailed description of a method of implementing the invention
    With reference to FIG. 1, the method of the invention is implemented by means of a device comprising a computer unit · 1 connected to a digitizing apparatus arranged to digitize a written document. L computer unit 1 is a computer which comprises at least one processor and a memory containing an image acquisition program and a program implementing the method of the invention. The processor is / arranged to execute: these: programs. The scanning device is for example a scanner 2 dedicated to the scanning of written documents (commonly called flatbed scanner), or else an image sensor of a communication terminal such as a smartphone 3 (more commonly referred to as its English name "smartphone") connectable to computer unit 1 via a network such as the Internet :. The scanner 2 is here controlled directly by the computer unit 1 to acquire the image of the document. As a variant, the scanner 2 can be connected to another computer unit which will control the acquisition of the image and will send the image to the computer unit 1 which will carry out image processing and character recognition proper . In the case of capture by the ordiphohe 3, the user controls the acquisition of the image of the written document directly from the ordiphone 3 then transmits this image; to the computer unit 1 so that the latter provides image processing and character recognition proper. The scanning device is in all cases arranged; to capture an image of the written document; having sufficient resolution to allow the extraction of alphanumeric characters which would be present in the image; and to recognize said characters.
    The written document here is more particularly a document; identity such as an identity card or passport.
    In Figure 2 is shown an image; 10 of this identity document. Image 10 has been captured; by the digitizing device. On this image: 10, we can see that · the Identity document includes a photograph of its holder and alphanumeric character fields, namely here a "Date" field 11 and a "City" field 12. Obviously ·, the identity document actually comprises; other fields of alphanumeric characters: - such as fields "Name", "First names", "Date of birth", "Place of birth", "Nationality", "Address", "End of validity date" - which were not represented here. In the following of; the description, the word 'characters' alone will be used to designate the alphanumeric characters. The identity document also includes security or decorative elements likely to interfere with; · written characters (not shown in Figure 2).
    The method of the invention implemented by the program executed by the computer unit 1 comprises the following steps (Ligure 3):
    - segmenting the image to identify therein 5 objects (step 110);
    - Define an enclosing box 20 around each object and make a first selection to select the enclosing boxes presumably containing a character: as a function of at least one · theoretical dimensional characteristic of an alphanumeric character (step 120);
    make a second selection comprising the application to each selected bounding box of shape descriptors and the implementation of a decision-making algorithm 15 to select, on the basis of the descriptors, the bounding boxes apparently containing a character (step 130 );
    group the bounding boxes according to positions relative bounding bdites (step 140);
    - perform a third selection by dividing each of these bounding boxes into a plurality of cells for each of which a texture descriptor is determined in the form of an oriented gradient histogram ·, the histograms then being concatenated and a prized decision algorithm · implemented to select, on the basis of the descriptors, the bounding boxes containing probably one character (step 15 0 );
    - improve a contrast of the image and detect contours · of objects present in the image so as to create a mask bringing out the characters (step 160);
    segment the image by applying the mask to the image to extract the objects visible therefrom through the mask (step 170);
    - perform character recognition on the finally selected bounding boxes (step 180).
    These steps will now be detailed.
    Step 110 here consists in applying to the image a sequential alternating filter which is a mathematical morphological filter. In practice, the program scans image 10 with a geometric window (commonly called a structuring element) which is in the form of a circle (but which could be rectangular or even linear or other) of 5 to 10 pixels in radius and eliminates all this; which fits entirely into said window (commonly called erosion) and expands any part of an object that does not fit fully into the window. Given the dimensions of the window, one character: will not fully fit inside the window and will therefore be dilated, the rest is necessarily noise and is eliminated. Preferably, several passes are made by increasing between each the dimensions of the window to gradually filter the noise of the image. As a variant, this step can be carried out by implementing an MSÉR type algorithm (from the English “Maximally stable 25 extremal regions”) or by filtering the image using a threshold corresponding to a theoretical intensity d 'a character (when the threshold is reached-, the object is considered to be a character; when the threshold is not: reached :, the object is not a character).
    At the end of this stage, the program therefore brought out objects (which could also be called: related components) which include alphanumeric characters as well as other objects including elements which are not :, such as security or decor elements. However, at this stage, a not insignificant part of these unwanted elements have been excluded.
    In step 120, on each of the: objects remaining in the image ;, the program applies a box; encompassing 20 (visible in FIG. 2) respecting several theoretical geometric criteria of the characters, namely: la; height, width and / or a ratio of dimensions (or ÆR of the English "aspect ratio"; height / width for example 10). If an object, and therefore: its bounding box: 20, have a: height and width (or a ratio thereof) corresponding to the theoretical ones of a character ,: it is an alphanumeric character. We can therefore) select objects: which can) correspond to 15 characters on the basis of geometric criteria.
    To automatically select the objects corresponding to alphanumeric characters) in step 130, the program ;: implements a decision-making algorithm (or more commonly called a classifier). On 20 each object; retained previously, several types of shape descriptors are determined, namely here;:;
    - the Fourier moments ·,
    - the moments of Krawtchouk.
    We recall that a moment is one; formula applied 25 on a pixel or a set of pixels making it possible to describe the structure which one tries to apprehend, namely here a character. Other descriptors could: be: used instead of or in addition to the Fourier moments and / or Krawtchouk moments. However, the combined 3: 0 use of these two: types of descriptors gives results: remarkable.
    Fourier moments are used in a classifier (here of SVM type from English << Support
    Vector Machine ”) to produce a first character / non-character output.
    The; Krawtchouk moments are used in a classifier (here again of SVM type); to produce a second character / non-character output.
    These two outputs are then · concatenated to form an input vector of a classifier - (here again of SVM type) providing a third output. This third output · is compared to a threshold to provide a binary decision: "character" or "no character". Preferably, to form the input vector, the first output and the second output are weighted for each object, for example according to the performance of the descriptors taking into account the type of background.
    Following this operation, an image is obtained containing; objects devoid of most of the possible stains and noise initially present · in the image, often due to the presence of security elements or decor of the document.
    In step 140, the program operates a grouping of characters in the form of one or more words or lines of text as a function of geometric criteria which, in addition to the height, the width and / or the ratio of dimensions AR , include the centroids (or barycenters) 25 of the bounding boxes 20; associated with each character.
    More specifically, the program detects if the · centroids are; aligned on the same line and Calculates · the distances separating the centroids of bounding boxes 20 associated with adjacent characters to determine 30 if they: belong to the same word. The grouped characters are associated in a collective bounding box.
    In step 150, the program examines the content of each collective bounding box: and eliminates those which do not seem to contain a text field. Indeed, during the phases described above, it may be that lines are unfortunately formed by grouping objects of which at least one is not a character ·. This 5 step therefore eliminates false positives.
    We know that different regions of text have distinct distributions of gradient orientations: the reason is that the gradients of high amplitude are generally perpendicular to the contours which form the characters. The program uses for this step a texture descriptor based on UH oriented gradient histogram or HOG (from the English "Histogram of oriented gradient") which is known in text recognition. Conventionally:
    - the area to be recognized is subdivided into NI lines and
    Do overall columns on the image f
    - a histogram is calculated for each of the NlxNc cells,
    - histograms - are concatenated with each other for the whole image.
    According to the method of the invention, the program is advantageously arranged to subdivide the bounding box 20 of each object into 3 rows and 1 column because this division makes it possible to significantly improve the decision "word" or "not word". Thus, on each of the three cells of each bounding box 20 containing a priori a character is calculated a histogram. The histograms are then concatenated: to each other then · introduced into a classifier (here again of type 30 SVM) to decide whether the collective bounding box: corresponds to text. Note that the division · is strongly dependent on the size of: characters. The bounding box: 20 in which we cut: must be: the size of each character (if the bounding box 20 of a character is 28 pixels x 28 pixels initially but the character occupies only 50% of it, we resize the box so that the character occupies all of it, then we do the cutting).
    -In step · 160, the program proceeds, in each collective encompassing, to a color analysis of the image - (two parts of the image before carrying out this step are shown in: figures · 4. a and 5. a): the objective here is to saturate the large differences in the image and to amplify the small differences by saturating the channels with: color (RGB, i.e. red, green, blue) to bring out the color of the characters (in the case of a black and white image, we will act on the gray levels). For this, the program performs a contrast enhancement which consists in locally adapting the contrast of the image by a lateral inhibition - difference of neighboring pixels - weighted by the Euclidean distance between the pixels. Only the strongest gradients are retained. Finally, the program also adapts the image in order to obtain a balance: overall white: (see the two parts of the image after step 160 in Figures 4.b and 5.b). This step improves the contrast and: corrects the color. Alternatively, a histogram equalization algorithm could have been used <25, but such an algorithm · produces artifacts and artificial colors in the backgrounds which may complicate further processing of the image.
    30 Step 170 aims to remove the background of the image in order to get rid of any background element therein such as - security elements · or decor, likely to disturb: later the recognition of characters :.
    The previous step improved the color of the image and saturated the: black characters. It is soon; then · easier to detect the outlines of characters. The method of the invention implemented by the program; uses for this purpose; a filter; contour detection and more; especially a Sobel filter.
    The image obtained as an output (Figure 5.c) is then used as a mask in a: segmentation approach by tree of connected components. In general, the trees of connected components associate: with an image: of gray level :, a descriptive data structure induced by an inclusion relation between the binary connected components obtained by the successive application of the lines of
    1: 5 level. The use of the mask. allows you to select in the tree only what concerns · characters ·. This selection: is carried out in an automatic manner so that the segmentation by tree of connected components can be carried out in an automatic manner, without human intervention, whereas, conventionally, the segmentation by tree of connected components implements a process interactive with · an operator '. The segmentation of a field by the method of the invention can thus be carried out much faster than with the conventional method. From tests carried out by the Applicants have shown that the segmentation by the process of the invention was faster in a ratio greater than 60 or even 70. Thus, the segmentation according to the invention makes it possible to reduce the calculation time.
    The character recognition performed by the program in step 180 can implement; no matter: which character recognition algorithm. More specifically, the pfbgramœe applies a segmentation and word recognition model which is based on a deep learning architecture based on a combination of convolutional neural networks: (CNN) and LS TM (CNN in English Convolutional Neural Network, LSTM in English Long Short-Term Memory). In the present case, the convolutional neural network gives particularly good results since the background of the image has been eliminated before its implementation. This elimination of the background decreases the rate of false positives during OCR; and in particular avoids the appearance of ghost characters, that is to say patterns from the background and / or security or decorative elements, these patterns having a shape close to that of a character and being recognized erroneously as being a character during OCR.
    Preferably, a multi-scale approach will be carried out as a variant. Indeed, the characters which are larger than · the window used during step 110 are often over-segmented. To avoid this drawback, the method according to the invention provides for carrying out steps 110 and 120 at different resolutions, the dimensions of the window remaining identical. In practice, the program proceeds to several passes: of scans · and decreases the resolution after each pass to eliminate each time; all objects which do not fully fit in the window but which have sizes 2; 5 smaller than that of a character ;. As; for example, the initial resolution is 2000x2000 pixels and five decreases are made; the resolution (the resolution is halved each time ;;). A number of five decreases represents a good compromise between efficiency and 30 calculation times.
    Note that the geometric criteria; relevant to the grouping of characters and the choice of different ones: parameters enabling effective detection of words to be selected; in order to have one; set of effective parameters for each type of image (depending on the wavelength range used for scanning: visible, IR and ÜV).
    Of course, the invention is not limited to the mode of implementation described but encompasses any variant coming within the scope of the invention as defined in the claims Attachments .
    In particular, the process has been described in its most efficient version whatever the digitizing device 10 used.
    for a scan by a flatbed scanner, the process of the invention may include only the following steps:
    - enhance image contrast;
    - detect the contours of objects present in the lice image to create a mask making the characters resound;
    segment the image by applying the mask to the image to extract the objects visible through it
    mask;
    - perform a: character recognition on the extracted objects.
    For scanning by smartphone, the method of the invention may include only: the following steps:
    - segment the image to identify objects therein;
    - define a bounding box around each object and and kill a first selection to select the bounding boxes apparently containing
    3 0 a character as a function of at least one theoretical dimensional characteristic of an alphanumeric character;
    make a second selection comprising the application to each selected bounding box of shape descriptors and the implementation of a decision-making algorithm to select, on the basis of the) descriptors, the bounding boxes supposedly containing a character;
    group the bounding boxes according to relative positions of the bounding boxes;
    - make a third selection by dividing each of these bounding boxes into a plurality of cells: for each of which is. determined a texture descriptor in the form of an oriented gradient histogram ::, the histograms then being concatenated and a decision-making algorithm being implemented to select, on the. base of descriptors, bounding boxes) supposedly containing a character ·;
    - perform character recognition on the finally selected bounding boxes :.
    In all cases., The multiscale approach is optional.
    It is possible to combine several classifiers. Or to use other classifiers than: those indicated ^. Preferably, each classif used to them) will be of a type 25 included in the) grouped as follows: SVM (from English “Support Vector Machine”), RVM (from English “Relevance Vector Machine”), K closer neighbors (or KNN), Random Forest. Note for example that: the RVM classifier allows a probabilistic interpretation 30 allowing to have fewer examples for the learning phase.
    It is possible to group by line or by word. One will take account for example of the type of document: thus on the identity documents of British origin, there are sometimes; between the letters of large spaces which leaves the background: very apparent: it is more efficient to perform a grouping; per word 'for 5 this type of document.
    For step 150, other divisions can be envisaged, in particular 1 column and 7 lines.
    Images can be processed in color or in; shades of grey. In grayscale ·, the use of the mask makes it possible to eliminate a large number: of parasitic elements.
    As a variant, several other segmentation solutions could have been envisaged, such as global or adaptive thresholding, a mixture of gaussians 15 or any other technique in order to efficiently isolate the characters of the image.
    Krawtchouk moments can be used alone or in combination with other types of moment and for example: form descriptors: based: also on 20 of the following moments: moments; de Fourier, de Legendre., de Zernike, de Hu and 'descriptors extracted: by a convolutional neural network of the type: LeNët. It will be noted that the moments of 'Krawtchouk become effective descriptors for the characters by using 25 polynomials of order: 9 while polynomials of order 16 are necessary for the moments of Legendre, 17 for the moments of Zernike and more than 30 for the moments: de Fourier.
    Θη note that the method of the invention is particularly well suited for processing documents having heterogeneous backgrounds. The process: can be implemented in the same way for the processing of documents with homogeneous backgrounds. It is also possible to provide a preliminary step to determine whether the background of the document is homogeneous and, if so, to pass the steps of contour detection and segmentation by mask. This segmentation is especially useful because it eliminates a large part of the background of the document which could alter the recognition of character. However, with a homogeneous background, this risk is limited. Another type of segmentation can possibly be envisaged.
    The device may have a structure different from that described. The image acquisition program can in particular be stored in a memory of the capture member in order to be: executed directly by the latter. The device and the capture device can be incorporated in the same device.
    权利要求:
    Claims (14)
    [1]
    1. Method for recognizing characters in an image of a document comprising at least one field
    5 alphanumeric, the method comprising the steps of:
    - segment the image to · identify objects in it;
    - define a bounding box around each object and make a first selection to
    10 selecting the bounding boxes supposedly containing a character · as a function of at least one theoretical dimensional characteristic of an alphanumeric character;
    make a second selection comprising the application to each selected bounding box of 15 shape descriptors and the implementation of a decision-making algorithm to select, on the basis of the descriptors, the enclosing boxes: supposedly containing a character;
    group the bounding boxes according to 20 relative positions of the bounding boxes;
    - make a third selection by dividing each of these bounding boxes into a plurality of cells for each of which a texture descriptor is determined in the form of a histogram of
    25 oriented gradient, lës : histograms then being concatenated and a decision-making algorithm implemented to select, on the basis of the descriptors, the bounding boxes supposedly containing a character;
    30 - perform a character recognition · on the finally selected bounding boxes.
    [2]
    2. Method according to claim 1, in which the shape descriptors are based at least on Krawtchouk moments.
    [3]
    3. Method according to claim 2, in which the form descriptors are also based on moments from the following: Fourier, Legendre, Zernike r de Hu moments and from descriptors extracted by a convolutional neural network of the LeNet type. .
    [4]
    4. The method of claim 2, wherein the second selection comprises r
    - 1 / application of shape descriptors based on Fourier moments and the implementation of a decision-making algorithm to obtain a first spell on,
    - the application of form descriptors based on Krawtchouk moments and the implementation of a decision-making algorithm to obtain a second output,
    - the weighting of outputs; to form an input vector of a decision-making algorithm having a third output compared to a threshold for deciding the existence of a character or not.
    [5]
    5. Method according to any one of the preceding claims, the dimensional characteristic used during the first selection is a ratio of dimensions.
    [6]
    6. Method according to any one of the preceding claims, in which the grouping of the bounding boxes is carried out by determining a barycenter of each 'bounding box, and by checking whether the barycenters are on the same line taking account of a spacing between barycentres.
    [7]
    7. Method according to any one of the preceding claims, in which the histogram is determined according to the HOG method with a division of the bounding box into three: rows and a column, ie three
    5 cells.
    [8]
    8. Method according to any one of the preceding claims, in which at least one of the decision-making algorithms is of a chosen type, from the following group: SVM, RVM, KNN or Random Forest.
    [9]
    9. Method according to any one of the preceding claims, in which the segmentation is carried out by scanning the image with a window having dimensions smaller than the theoretical dimensions of a character, by eliminating any object entering entirely into the window.
    15 and by expanding any part of an object which does not fully fit into the window.
    10. The method of claim 8, wherein the segmentation is performed at multiple resolutions; the method comprising the step, for each resolution,
    20 to eliminate any object that does not fit entirely in the window, the window remaining the same size.
    [10]
    11. The method of claim 10, wherein the number of resolutions is at most five.
    [11]
    12. Method according to any one of
    25 previous claims ·, ’wherein character recognition is performed by a neural network.
    [12]
    13 :. Method according to the preceding claim, in which the neural network is of the short and long term memory convolution type.
    30
    [13]
    14- Method according to any one of the preceding claims /, comprising, prior to character recognition, the steps of:
    - reinforce a contrast of the image to bring out the characters present in the image detect contours of objects present in the image to create a mask bringing out the 5 characters;
    segment the image using a tree with connected components and applying the mask to it so as to extract the characters from the image.
    [14]
    15. Character recognition device
    10 comprising a computer unit (1) provided with means for its connection to a digitizing apparatus arranged to perform a digitization of a written document, characterized in that the computer unit (!) Comprises at least one processor and a memory containing a
    15 program implementing the method according to any one of the preceding claims.
    类似技术:
    公开号 | 公开日 | 专利标题
    EP3570212A1|2019-11-20|Character recognition method
    CA3043090C|2021-07-27|Character recognition process
    EP1298588B1|2011-08-03|Image processing system for the automatic extraction of semantic elements
    FR3059804B1|2019-08-02|IMAGE PROCESSING SYSTEM
    WO2004013802A2|2004-02-12|Method and system for automatically locating text areas in an image
    FR2795205A1|2000-12-22|METHOD FOR BINARIZING DIGITAL IMAGES WITH MULTIPLE GRAY LEVELS
    EP0910035A1|1999-04-21|Method for automatic extraction on printed text or handwriting on a background, in a multi-level digital image
    EP3506170A2|2019-07-03|Method for forming a neural network for recognising a sequence of characters and associated recognition method
    FR2861524A1|2005-04-29|Method for detecting orientation of image taken by digital camera, involves detecting lines in image, calculating attributes that characterize lines, for each detected line, and detecting orientation of image according to attributes
    BE1021013B1|2014-12-11|METHOD AND SYSTEM FOR IMPROVING THE QUALITY OF COLOR IMAGES
    EP3832535A1|2021-06-09|Method for detecting at least one visible element of interest in an input image by means of a convolutional neural network
    EP1684210B1|2009-08-26|Red-eye detection based on skin region detection
    FR3112228A1|2022-01-07|Device and method for generating a mask of the silhouette of the profile of a structure
    FR3105529A1|2021-06-25|A method of segmenting an input image representing a document containing structured information
    FR3091610A1|2020-07-10|Method for processing digital images
    FR3095286A1|2020-10-23|Process for image processing of an identity document
    WO2021144427A1|2021-07-22|Method for processing a candidate image
    EP3931751A1|2022-01-05|Computer-assisted image-processing method for reading a marking on a surface of an animal-skin support
    FR3049094A1|2017-09-22|METHOD OF IDENTIFYING A SET OF ROW BOOKS TO AUTOMATICALLY ESTABLISH THEIR LIST
    FR3057690A1|2018-04-20|METHOD OF COMPLETING AN IMAGE OF A SCENE UTILIZING THE SPECTRAL CHARACTERISTICS OF THE SCENE
    FR3030835A1|2016-06-24|METHOD FOR CLASSIFYING A DOCUMENT FROM ITS VISUAL ASPECT
    FR2910670A1|2008-06-27|Natural image binarizing method, involves constructing binary text image by neural architecture having heterogeneous artificial neurons layers in iterative manner, and binarizing image comprising text by image construction operation
    同族专利:
    公开号 | 公开日
    AU2019203344A1|2019-12-05|
    AU2019203344B2|2020-03-19|
    US20190354756A1|2019-11-21|
    US10885326B2|2021-01-05|
    FR3081245B1|2020-06-19|
    CA3043090A1|2019-11-17|
    EP3570213A1|2019-11-20|
    CA3043090C|2021-07-27|
    引用文献:
    公开号 | 申请日 | 公开日 | 申请人 | 专利标题
    US20060045322A1|2004-08-26|2006-03-02|Ian Clarke|Method and system for recognizing a candidate character in a captured image|
    US9235757B1|2014-07-24|2016-01-12|Amazon Technologies, Inc.|Fast text detection|
    WO2017058252A1|2015-10-02|2017-04-06|Hewlett-Packard Development Company, L.P.|Detecting document objects|
    US7650041B2|2006-02-24|2010-01-19|Symbol Technologies, Inc.|System and method for optical character recognition in an image|
    US9165187B2|2012-01-12|2015-10-20|Kofax, Inc.|Systems and methods for mobile image capture and processing|
    US9760789B2|2012-04-17|2017-09-12|Conduent Business Services, Llc|Robust cropping of license plate images|
    US9830508B1|2015-01-30|2017-11-28|Quest Consultants LLC|Systems and methods of extracting text from a digital image|
    FR3045884B1|2015-12-18|2017-12-08|Morpho|METHOD FOR DETECTING FRAUD TO AUTHENTICATE A FINGER|
    FR3048540B1|2016-03-04|2019-09-13|Idemia Identity And Security|IMAGE ENHANCEMENT METHOD APPLICABLE TO DIGITAL IMPRESSION IMAGES|
    FR3081244B1|2018-05-17|2020-05-29|Idemia Identity & Security France|CHARACTER RECOGNITION PROCESS|
    US10719700B1|2018-06-13|2020-07-21|Lendingclub Corporation|Optical character recognition improvement based on inline location determination|
    US11055527B2|2019-02-01|2021-07-06|Intuit Inc.|System and method for information extraction with character level features|US11087163B2|2019-11-01|2021-08-10|Vannevar Labs, Inc.|Neural network-based optical character recognition|
    WO2021243294A1|2020-05-29|2021-12-02|Camelot Uk Bidco Limited|Method, non-transitory computer-readable storage medium, and apparatus for searching an image database|
    CN113807326B|2021-11-17|2022-02-25|航天宏康智能科技有限公司|System table character recognition method and device|
    法律状态:
    2019-04-18| PLFP| Fee payment|Year of fee payment: 2 |
    2019-11-22| PLSC| Publication of the preliminary search report|Effective date: 20191122 |
    2020-04-22| PLFP| Fee payment|Year of fee payment: 3 |
    2021-04-21| PLFP| Fee payment|Year of fee payment: 4 |
    优先权:
    申请号 | 申请日 | 专利标题
    FR1854139A|FR3081245B1|2018-05-17|2018-05-17|CHARACTER RECOGNITION PROCESS|
    FR1854139|2018-05-17|FR1854139A| FR3081245B1|2018-05-17|2018-05-17|CHARACTER RECOGNITION PROCESS|
    CA3043090A| CA3043090C|2018-05-17|2019-05-10|Character recognition process|
    EP19174230.3A| EP3570213A1|2018-05-17|2019-05-13|Character recognition method|
    AU2019203344A| AU2019203344B2|2018-05-17|2019-05-13|Character recognition method|
    US16/415,574| US10885326B2|2018-05-17|2019-05-17|Character recognition method|
    [返回顶部]