CA2502529A1 - Model of documents and method for automatically classifying a document - Google Patents
Model of documents and method for automatically classifying a document Download PDFInfo
- Publication number
- CA2502529A1 CA2502529A1 CA002502529A CA2502529A CA2502529A1 CA 2502529 A1 CA2502529 A1 CA 2502529A1 CA 002502529 A CA002502529 A CA 002502529A CA 2502529 A CA2502529 A CA 2502529A CA 2502529 A1 CA2502529 A1 CA 2502529A1
- Authority
- CA
- Canada
- Prior art keywords
- document
- documents
- identifying tag
- class
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Abandoned
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V30/00—Character recognition; Recognising digital ink; Document-oriented image-based pattern recognition
- G06V30/40—Document-oriented image-based pattern recognition
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/20—Image preprocessing
- G06V10/22—Image preprocessing by selection of a specific region containing or referencing a pattern; Locating or processing of specific regions to guide the detection or recognition
Abstract
A method for automatically creating a model of documents representing a class of documents, the method comprising the steps of:
a) providing (100) a plurality of documents separated into different classes;
b) determining (200) at least one potential identifying tag within one document (8) belonging to one class (A), said identifying tag (1, 2) being defined by at least its graphical content, size and location within the document;
c) checking whether (300) said potential identifying tag is included within at least a document of said one class (A);
d) if said step c) is positive, selecting (330) said identifying tag;
d') if step c) is negative, rejecting (340) said potential identifying tag and repeating steps b) and c);
e) creating a model of documents (10) related to said one class (A) including said selected identifying tag (1, 2).
a) providing (100) a plurality of documents separated into different classes;
b) determining (200) at least one potential identifying tag within one document (8) belonging to one class (A), said identifying tag (1, 2) being defined by at least its graphical content, size and location within the document;
c) checking whether (300) said potential identifying tag is included within at least a document of said one class (A);
d) if said step c) is positive, selecting (330) said identifying tag;
d') if step c) is negative, rejecting (340) said potential identifying tag and repeating steps b) and c);
e) creating a model of documents (10) related to said one class (A) including said selected identifying tag (1, 2).
Description
MODEL OF DOCUMENTS AND METHOD FOR AUTOMATICALLY
CLASSIFYING A DOCUMENT
The invention relates to the field of document image processing, and more specifically to document image classification. We refer to document image as a document that has been digitalized by any means.
In various applications, it is desirable to classify documents by their type, e.g., business letters, invoice, fax cover sheet, and by their origin, e.g., customer, subscriber, etc.
Obviously, documents can be classified as belonging to one identifiable class.
We define a class as being a set of documents of a given type, each document including a structure and/or contents similar to the other documents of the class and different from the documents of any other class. Any pair of documents taken from the same class should have at least some regions similar to each other and any pair of documents taken from two different classes should have at least some regions dissimilar to each other.
United States Patent 6,542,635 discloses a method wherein a document to be classified is segmented into blocks of data, for instance by using a pattern or optical character recognition (ORC) applied to part or the entire document. A vector of characteristics is created, representing a segment or the entire imaged document. A
classification algorithm is then applied to the vector of characteristics to determine to which class the document belongs.
Various classification algorithms are known, such as the K-mean method, the fuzzy C-mean method, and neuronal network based approaches. However, automatic classification of document images requires high processing capacity.
Some known methods perform document classification further to a learning step. Models are created to define different classes of documents. Documents to be classified are then assigned to a class by comparing the document to be classified with the different models. Such models of documents are commonly created by an operator and such supervised operation is time consuming and expensive.
According to the invention a new method of classification is provided that can increase the rate of document processing and suppress classification system operator intervention.
Moreover, the method according to the invention remains reliable and efficient when the number of classes increases.
According to the invention, models of documents are created starting from a sample of classified documents. Such models are representative of the contents of all the documents of a given class while including less data to be processed.
\1HIRSCH6~BREVETS\Brevets~20400~20497EP.doc- 14/03/05- 15:03 - 1/14 When such models of documents are created, an incoming document to be classified is compared to the models and assigned to a class or rejected.
Therefore, the invention provides a specific way to model a class of documents and a method for automatically creating such models of documents.
Furthermore, the invention provides a method for assigning a class to a document using models of documents according to the invention.
The methods of the present invention are preferably computer-implemented.
In particular, the invention concerns a model of documents including at least one identifying tag, each identifying tag being defined by at least its graphical content, size and location within the document.
According to a feature, the object to surface ratio of the graphical content of the identifying tag is substantially equal to 0,5.
According to another feature, the size of the identifying tag is less or equal to 5% of the document size.
The invention also concerns a method for automatically creating a model of documents representing a class of documents, the method comprising the steps of a) providing a plurality of documents separated into different classes;
b) determining at least one potential identifying tag within one document belonging to one class, said identifying tag being defined by at least its graphical content, size and location within the document ;
c) checking whether said potential identifying tag is included within at least a document of said one class;
d) if said step c) is positive, selecting said identifying tag;
d') if step c) is negative, rejecting said potential identifying tag and repeating steps b) and c);
e) creating a model of documents related to said one class including said selected identifying tag.
The method according to the invention may further comprise the steps of-.
fj checking whether said potential identifying tag is included within at least a document of another class;
g) if step f) is negative, selecting said identifying tag;
g') if step fj is positive, rejecting said potential identifying tag and repeating steps b) and c).
The method is repeated for each class of documents provided.
According to one feature, step b) comprises the steps of b~) selecting a working window in said one document;
bz) calculating an object to surface ratio for said working window;
\\HIRSCH6\BREV ETS\Breve1s120400\20497EP.doc - 14103/05 - 15:03 - 7l14 b3) if said ratio lies within predetermined values, selecting the region of said working window as a potential identifying tag.
According to one feature, steps b~) to b3) are repeated by shifting the working window to scan the document. A given portion of the document may be excluded from the searching of an identifying tag.
According to one feature, steps c) and fJ comprise the step of calculating a correlation function between the potential identifying tag and a region of said document corresponding to the size and location of said identifying tag.
The invention also concerns a computer system for creating a model of documents according to the method of the invention.
The invention also concerns a method for automatically classifying a document comprising the steps of a') providing a plurality of models of documents according to the invention, each model of documents being related to a given class;
b') comparing a document to be classified with at least one model of documents;
c') if said document to be classified matches with said one model of documents, assigning said document into said corresponding class;
d') if said document to be classified doesn't match with any model of documents, rejecting said document.
According to one feature, step b') includes the step of calculating a correlation function between at least one identifying tag of the model of documents and a region of the document to be classified corresponding to the size and location of said identifying tag.
According to one feature, step c') includes the steps of - comparing the correlation function results for each model of documents provided at step a' );
- assigning the document to be classified to the class corresponding to the model of documents providing the correlation result closest to a predetermined value.
The invention also concerns a computer system for classifying a document according to the method of the invention.
Other characteristics and advantages of the invention will appear on reading the following detailed description of the embodiments of the invention, given solely as an example, and with reference to the drawings which show - figure 1 a is a schematic view of a document to be classified;
\~HIRSCH6\BREVETS~Brevets~20400~20497EP.doc - 14/03/05 - I 503 - 3/ t4 - figure lb is a schematic view of a model representing the class of the document of figure 1 a;
- figure 2 is a flowchart of the method for automatically creating a model of documents according to the invention;
S - figure 3 is a flowchart of the method for automatically classifying a document according to the invention.
The invention proposes creating models of documents to represent classes of documents and that can be further used for document classification.
According to the invention, a model of documents includes at least one identifying tag or marker. We refer to an identifying tag as being a region of document representing said document. An identifying tag is defined by at least its graphical content, its size and location within the document.
Figures la and lb show respectively a document 8 and the corresponding model of documents 10 according to the invention.
As shown on figure lb, the model 10 includes a plurality of identifying tags 1, 2. However, one identifying tag may be enough to define a model of documents and up to several dozens identifying tags associated to one model can be handled without too much slowing down processing. Preferably, an average of 5 to 10 identifying tags is compliant to define most models of documents.
In the example of figure 1, the identifying tags 1, 2 respectively represent the originating company of the document 8, with its name, address and possibly with a logo, and the type of document 8 as being an invoice. The model 10 should therefore be representative of any document having the same originating company and being an invoice.
An identifying tag has a predetermined size, preferably limited to up to 5% of the document size, and preferably less than 1% of the document size. An identifying tag is located in a specific region of the document, e.g., at the top left, on a bottom line, etc. Defining several small sized tags spread over a given portion of document is preferred to a single bigger tag located on the same portion of document.
The identifying tag is also defined by its graphical contents. A region of the document selected as an identifying tag preferably has an object-to-surface ratio substantially equal to 0,5. This ratio ensures that the identifying tag includes representative data that can be computed to further compare documents.
As each document is digitalized, an identifying tag is therefore a set of pixels that can be processed independently from the whole document.
Each identifying tag 1, 2 has a given graphical content D1, D2, which is data related to the values of the pixels in the specific region of the tag. Each tag 1, 2 also \1HIRSCH6\BREVETS\Brcvets\20400\20497EP.doc - 14/03/05 - 15:03 - 4114 has a given size, S1, S2, which is the number of lines and columns of pixels in the specific region of the tag. Each tag also has a given location within the document, for instance defined by the upper left coordinates, (xl, yl), (x2, y2) of the specific region of the tag.
CLASSIFYING A DOCUMENT
The invention relates to the field of document image processing, and more specifically to document image classification. We refer to document image as a document that has been digitalized by any means.
In various applications, it is desirable to classify documents by their type, e.g., business letters, invoice, fax cover sheet, and by their origin, e.g., customer, subscriber, etc.
Obviously, documents can be classified as belonging to one identifiable class.
We define a class as being a set of documents of a given type, each document including a structure and/or contents similar to the other documents of the class and different from the documents of any other class. Any pair of documents taken from the same class should have at least some regions similar to each other and any pair of documents taken from two different classes should have at least some regions dissimilar to each other.
United States Patent 6,542,635 discloses a method wherein a document to be classified is segmented into blocks of data, for instance by using a pattern or optical character recognition (ORC) applied to part or the entire document. A vector of characteristics is created, representing a segment or the entire imaged document. A
classification algorithm is then applied to the vector of characteristics to determine to which class the document belongs.
Various classification algorithms are known, such as the K-mean method, the fuzzy C-mean method, and neuronal network based approaches. However, automatic classification of document images requires high processing capacity.
Some known methods perform document classification further to a learning step. Models are created to define different classes of documents. Documents to be classified are then assigned to a class by comparing the document to be classified with the different models. Such models of documents are commonly created by an operator and such supervised operation is time consuming and expensive.
According to the invention a new method of classification is provided that can increase the rate of document processing and suppress classification system operator intervention.
Moreover, the method according to the invention remains reliable and efficient when the number of classes increases.
According to the invention, models of documents are created starting from a sample of classified documents. Such models are representative of the contents of all the documents of a given class while including less data to be processed.
\1HIRSCH6~BREVETS\Brevets~20400~20497EP.doc- 14/03/05- 15:03 - 1/14 When such models of documents are created, an incoming document to be classified is compared to the models and assigned to a class or rejected.
Therefore, the invention provides a specific way to model a class of documents and a method for automatically creating such models of documents.
Furthermore, the invention provides a method for assigning a class to a document using models of documents according to the invention.
The methods of the present invention are preferably computer-implemented.
In particular, the invention concerns a model of documents including at least one identifying tag, each identifying tag being defined by at least its graphical content, size and location within the document.
According to a feature, the object to surface ratio of the graphical content of the identifying tag is substantially equal to 0,5.
According to another feature, the size of the identifying tag is less or equal to 5% of the document size.
The invention also concerns a method for automatically creating a model of documents representing a class of documents, the method comprising the steps of a) providing a plurality of documents separated into different classes;
b) determining at least one potential identifying tag within one document belonging to one class, said identifying tag being defined by at least its graphical content, size and location within the document ;
c) checking whether said potential identifying tag is included within at least a document of said one class;
d) if said step c) is positive, selecting said identifying tag;
d') if step c) is negative, rejecting said potential identifying tag and repeating steps b) and c);
e) creating a model of documents related to said one class including said selected identifying tag.
The method according to the invention may further comprise the steps of-.
fj checking whether said potential identifying tag is included within at least a document of another class;
g) if step f) is negative, selecting said identifying tag;
g') if step fj is positive, rejecting said potential identifying tag and repeating steps b) and c).
The method is repeated for each class of documents provided.
According to one feature, step b) comprises the steps of b~) selecting a working window in said one document;
bz) calculating an object to surface ratio for said working window;
\\HIRSCH6\BREV ETS\Breve1s120400\20497EP.doc - 14103/05 - 15:03 - 7l14 b3) if said ratio lies within predetermined values, selecting the region of said working window as a potential identifying tag.
According to one feature, steps b~) to b3) are repeated by shifting the working window to scan the document. A given portion of the document may be excluded from the searching of an identifying tag.
According to one feature, steps c) and fJ comprise the step of calculating a correlation function between the potential identifying tag and a region of said document corresponding to the size and location of said identifying tag.
The invention also concerns a computer system for creating a model of documents according to the method of the invention.
The invention also concerns a method for automatically classifying a document comprising the steps of a') providing a plurality of models of documents according to the invention, each model of documents being related to a given class;
b') comparing a document to be classified with at least one model of documents;
c') if said document to be classified matches with said one model of documents, assigning said document into said corresponding class;
d') if said document to be classified doesn't match with any model of documents, rejecting said document.
According to one feature, step b') includes the step of calculating a correlation function between at least one identifying tag of the model of documents and a region of the document to be classified corresponding to the size and location of said identifying tag.
According to one feature, step c') includes the steps of - comparing the correlation function results for each model of documents provided at step a' );
- assigning the document to be classified to the class corresponding to the model of documents providing the correlation result closest to a predetermined value.
The invention also concerns a computer system for classifying a document according to the method of the invention.
Other characteristics and advantages of the invention will appear on reading the following detailed description of the embodiments of the invention, given solely as an example, and with reference to the drawings which show - figure 1 a is a schematic view of a document to be classified;
\~HIRSCH6\BREVETS~Brevets~20400~20497EP.doc - 14/03/05 - I 503 - 3/ t4 - figure lb is a schematic view of a model representing the class of the document of figure 1 a;
- figure 2 is a flowchart of the method for automatically creating a model of documents according to the invention;
S - figure 3 is a flowchart of the method for automatically classifying a document according to the invention.
The invention proposes creating models of documents to represent classes of documents and that can be further used for document classification.
According to the invention, a model of documents includes at least one identifying tag or marker. We refer to an identifying tag as being a region of document representing said document. An identifying tag is defined by at least its graphical content, its size and location within the document.
Figures la and lb show respectively a document 8 and the corresponding model of documents 10 according to the invention.
As shown on figure lb, the model 10 includes a plurality of identifying tags 1, 2. However, one identifying tag may be enough to define a model of documents and up to several dozens identifying tags associated to one model can be handled without too much slowing down processing. Preferably, an average of 5 to 10 identifying tags is compliant to define most models of documents.
In the example of figure 1, the identifying tags 1, 2 respectively represent the originating company of the document 8, with its name, address and possibly with a logo, and the type of document 8 as being an invoice. The model 10 should therefore be representative of any document having the same originating company and being an invoice.
An identifying tag has a predetermined size, preferably limited to up to 5% of the document size, and preferably less than 1% of the document size. An identifying tag is located in a specific region of the document, e.g., at the top left, on a bottom line, etc. Defining several small sized tags spread over a given portion of document is preferred to a single bigger tag located on the same portion of document.
The identifying tag is also defined by its graphical contents. A region of the document selected as an identifying tag preferably has an object-to-surface ratio substantially equal to 0,5. This ratio ensures that the identifying tag includes representative data that can be computed to further compare documents.
As each document is digitalized, an identifying tag is therefore a set of pixels that can be processed independently from the whole document.
Each identifying tag 1, 2 has a given graphical content D1, D2, which is data related to the values of the pixels in the specific region of the tag. Each tag 1, 2 also \1HIRSCH6\BREVETS\Brcvets\20400\20497EP.doc - 14/03/05 - 15:03 - 4114 has a given size, S1, S2, which is the number of lines and columns of pixels in the specific region of the tag. Each tag also has a given location within the document, for instance defined by the upper left coordinates, (xl, yl), (x2, y2) of the specific region of the tag.
5 A model of documents according to the invention is defined by one or more identifying tags as just described.
Figure 2 illustrates a flow chart of the method for automatically creating a model of documents according to the invention.
The creation of models of documents is conducted starting from a plurality of classified documents. The object of the method according to the invention is to define a model for each class of the classification, such model being suitable for later use to classify subsequent documents.
The plurality of classified documents are organised in a set of different classes, each class being representative of a given type of document, according to the definitions given previously. The classified documents may come from any classifier using any known method of classification. The classified documents may be composed of a sample of at least one document, but preferably at least 2 documents per class to initiate the method for automatically creating a model according to the invention. The more documents we use to create a model of documents, the better the identifying tags will be representative of the class of these documents.
Referring to figure 2, the first step 100 of the method according to the invention is to provide a plurality of classified documents belonging to different classes.
A second series of steps 200 is to determine at least one potential identifying tag within one document belonging to one class.
One document is picked up (step 210) from one class A. A working window is selected (step 220) in said one document. The working window can be set with a predetermined size S1, for instance 10*10 pixels, which may possibly correspond to some relevant part of a company logo, name or address. The size of the working window should be big enough to contain relevant information and small enough to avoid slowing down processing. The working window is first located on a predetermined portion of document, for instance, upper left corner.
A ratio is then calculated (step 230) over said working window, representing the object surface divided by the window surface. For instance, pixels of the background (possibly white) are assigned a 0 value and the pixels representing patterns, letters or any other sign appearing on the background, for instance black letters and colour signs, are assigned the value 1.
\UiIRSCH6\BREVETSiBreveut20400\20497EP.doc - f4103/OS - 15:03 - 5/14 When the calculated ratio is substantially equal to 0,5; or when the value ratio-0,5 is minimized, this constitutes an indication that some information is located within the region where the working window is located. Such a region is then selected to be a potential identifying tag (step 240) with specific content data D, size S and position (x, y).
The working window is shifted to scan (step 250) the document to search for different identifying tags, by repeating the step of ratio computation on different portions of the document. A portion of the document may be excluded from the scan to prevent the selection of a tag that is known to be of no interest in the classification.
For instance, a bottom line to prevent selection of useless information of an endpaper page, or a region contiguous to an already selected potential tag.
A subsequent scan can be eventually conducted with another working window having a different size S2. Therefore, a plurality of potential identifying tags 1, 2 can be selected for one document, each tag having a specific content, i.e. data D1, D2, a specific size, S1, S2 and a specific position within the document (xl, yl), (x2, y2)_ After potential identifying tags have been selected in a first document according to the series of steps 200, a further series of steps 300 is performed to check whether each said potential identifying tag is representative of most documents of the class A.
Another document, from the same class A as said first document, is picked up (step 310). The region corresponding to the position and size of a potential identifying tag is selected on said other document and a correlation function is computed (step 320) between the potential identifying tag data D and the data of said other document region.
The correlation computation can be expressed for each tag of a model of documents as:
Cr g =1- ( P>>-P'>> ~ ~
With n: number of x coordinates pixels m: number of y coordinates pixels i: incrementation up to value n j: incrementation up to value m P;~: pixel data of the potential identifying tag P';~: pixel data of the corresponding region in the other document The correlation function therefore expresses the similarity between the potential identifying tag and the corresponding region in another document of the same class. If Cog is substantially equal to l, the similarity is high and the potential tag can be considered as being representative of the class A and preselected (step \\HIRSCH6\BREVETSIBrcvets120400\20497EP-dot - 14/03/05 - 15:03 - 6114 330). If not, the potential identifying tag must be rejected (step 340) as not being representative of the class.
The correlation may be refined by computing the function with slight shifts of pixels p' in both coordinate directions x, y and retaining the maximum value.
A
potential identifying tag may be present on another document with a small shift and must be detected anyway.
Such correlation can be conducted for some or each potential tag selected in the first document of the first class and for some or each other document of the same class A.
It will be understood that the routine can be stopped as soon as one match misses. In other words, as soon as one potential identifying tag selected in the first document does not match with one other document of the same class A, said potential identifying tag is rejected (step 340). In one embodiment, the routine can also be stopped as soon as one potential tag is included in one other document of the same class.
If the potential identifying tag matches with at least one other document of class A, said tag can be preselected (step 330). According to embodiments, the check of the potential identifying tag matching with other documents of class A can be conducted on some documents or on all documents of said class A. If class A
only contains one document, all potential tags identified at step 240 will be preselected (step 330).
When an identifying tag is preselected after the series of steps 300, a further series of steps 400 can be conducted to check whether said preselected tag does not match with the documents of other classes, that is, to verify that said preselected tag is discriminant vis-a-vis other classes.
A document of a different class B is picked up (step 410) and a correlation function is computed (step 420) based on the same mathematical expression as set before. If Cog is then substantially equal to 0, the similarity between the preselected identifying tag and a corresponding region of the document of class B is low and said identifying tag can be considered as being discriminant over class B.
Such discrimination control can be conducted over all other classes of documents or stopped as soon as one class B is not represented by said preselected tag.
At least one preselected identifying tags that has not been found in another document of another class is then selected (step 430) as being an identifying tag representing class A otherwise the preselected identifying tag is rejected (step 440).
A model of documents 10 can further be created including said at least one selected identifying tags (i, 2) to represent class A.
\WIRSCH6\BREVETS\Brevets\20400\20497EP,doc - 11/03!05 - 15:03 - 7114 g It must be understood that the steps of selecting potential identifying tags can be mixed with the steps of preselection 300 and selection 400 of an identifying tag. That is to say, when a potential identifying tag is found in one document 8 of one class A, the steps of searching for said identifying tag in other documents of the same S class A and checking that said identifying tag does not match with another document of another class B, can be conducted before another potential identifying tag is found in first document 8.
All the steps of the method for automatically creating a model of documents, as shown in figure 2, can be software implemented.
The invention therefore refers to a computer system for creating a model of documents representing a class of documents comprising computer-implemented means for accessing a plurality of documents separated into different classes;
for determining at least one potential identifying tag within one document belonging to one class; for checking whether said potential identifying tag is included within at least a document of said one class ; for selecting said identifying tag as representing said one class if previous checking is positive; for rejecting said potential identifying tag if previous checking is negative; for creating a model of documents related to said one class including said selected identifying tag.
The computer system may also include computer-implemented means for checking whether said identifying tag is included within at least a document of another class; for selecting said identifying tag as representing said one class if previous checking is negative; for rejecting said preselected identifying tag if previous checking is positive.
The models of documents according to the invention are automatically created with no need for an operator to intervene in the process.
Applicant has run a software according to the method of the invention to create 223 classes starting from an image database of 5 000 digitalized invoices.
Each class includes about 20 documents. This software was run with a Pentium 4 to 2 Giga Hertz and took about 1h30 to create the models.
According to the invention, a set of models of documents is created respectively representing different classes.
Subsequently, when a document to be classified comes in, the models of documents can be used to classify said document, which means to assign a class to said document or to reject said document if not belonging to any known classes. The invention concerns therefore also a method for automatically classifying a document.
Figure 3 is a flow chart illustrating the classification method according to the invention.
\\HIRSCH6\BREVETS\Brevec~\20400\20497EP.doc - 14!03/05 - 1$:03 - &14 There is provided a plurality of models of documents according to the invention (step 500), each model of documents being related to a class.
The document to be classified is compared with a first model of documents refernng to a first class A.
At least one region of the document to be classified is selected, said region corresponding to the size and location of at least one identifying tag of said f rst model of documents. A correlation function is computed (step 510) between said at least one identifying tag and said region of the document to be classified.
The correlation function between the document to be classified and the model of a given class is based on the following expression:
C~ms = ( Crag) ~ Ntag With Crag: the correlation between a tag of the model of documents and a corresponding region of the document to be classified based on the expression set before, and N~g: the number of tags of the model of documents. Nag can be equal or set to 1 according to one embodiment.
If C~~~S is substantially equal to 1, said document matches with said model of documents and can be assigned to the corresponding class (step 520). If C~~~S
is substantially equal to 0, said document does not match with said model of documents.
A subsequent model of documents of another class is picked up (step 530) and the correlation C~nss is calculated (step 540) for said other class.
According to embodiments, the correlation C~~~S can be computed for all defined classes and the document to be classified will be assigned to the class for which C~,~S has the closest value to 1; or the routine will be stopped as soon as one computed value of C~,~S exceeds a threshold value.
If said document to be classified does not match with any model of documents, that is to say if C~tass is smaller than a threshold value substantially close to 0 for all models of documents, said document must be rejected (step 550) as not being suitable for classification according to the method.
All the steps of the method for automatically classifying a document, as shown in figure 3, can be software implemented.
The invention therefore refers to a computer system for classifying a document comprising computer-implemented software means for providing a plurality of models of documents, each model of documents being related to a class; for comparing a document to be classified with one model of documents; for assigning said document to a given class, when said document matches with the corresponding \\HIRSCH6\BREVETS\Brevets\20400\20497EP.doc - 14/03105 - 15:03 - 9l14 one model of documents; for rejecting said document to be classified, if said document to be classified does not match with any model of documents.
The classification method according to the invention increases the rate of automatic document image processing. The document to be classified is solely 5 compared with models of documents by computing data over specific regions of document images corresponding to tags location.
Applicant has run a software according to the method of the invention to classify a document using 223 models created according to the method of the invention. This software was run using a Pentium 4 to 2 Giga Hertz and it took about 10 66,6ms to assign a document to a given class (classification rate is about 15 image documents per second).
1\HIRSCH6\BREVETS\Brevas\20400\20497EP.doc - 14/03/05 - 15.03 - 10/14
Figure 2 illustrates a flow chart of the method for automatically creating a model of documents according to the invention.
The creation of models of documents is conducted starting from a plurality of classified documents. The object of the method according to the invention is to define a model for each class of the classification, such model being suitable for later use to classify subsequent documents.
The plurality of classified documents are organised in a set of different classes, each class being representative of a given type of document, according to the definitions given previously. The classified documents may come from any classifier using any known method of classification. The classified documents may be composed of a sample of at least one document, but preferably at least 2 documents per class to initiate the method for automatically creating a model according to the invention. The more documents we use to create a model of documents, the better the identifying tags will be representative of the class of these documents.
Referring to figure 2, the first step 100 of the method according to the invention is to provide a plurality of classified documents belonging to different classes.
A second series of steps 200 is to determine at least one potential identifying tag within one document belonging to one class.
One document is picked up (step 210) from one class A. A working window is selected (step 220) in said one document. The working window can be set with a predetermined size S1, for instance 10*10 pixels, which may possibly correspond to some relevant part of a company logo, name or address. The size of the working window should be big enough to contain relevant information and small enough to avoid slowing down processing. The working window is first located on a predetermined portion of document, for instance, upper left corner.
A ratio is then calculated (step 230) over said working window, representing the object surface divided by the window surface. For instance, pixels of the background (possibly white) are assigned a 0 value and the pixels representing patterns, letters or any other sign appearing on the background, for instance black letters and colour signs, are assigned the value 1.
\UiIRSCH6\BREVETSiBreveut20400\20497EP.doc - f4103/OS - 15:03 - 5/14 When the calculated ratio is substantially equal to 0,5; or when the value ratio-0,5 is minimized, this constitutes an indication that some information is located within the region where the working window is located. Such a region is then selected to be a potential identifying tag (step 240) with specific content data D, size S and position (x, y).
The working window is shifted to scan (step 250) the document to search for different identifying tags, by repeating the step of ratio computation on different portions of the document. A portion of the document may be excluded from the scan to prevent the selection of a tag that is known to be of no interest in the classification.
For instance, a bottom line to prevent selection of useless information of an endpaper page, or a region contiguous to an already selected potential tag.
A subsequent scan can be eventually conducted with another working window having a different size S2. Therefore, a plurality of potential identifying tags 1, 2 can be selected for one document, each tag having a specific content, i.e. data D1, D2, a specific size, S1, S2 and a specific position within the document (xl, yl), (x2, y2)_ After potential identifying tags have been selected in a first document according to the series of steps 200, a further series of steps 300 is performed to check whether each said potential identifying tag is representative of most documents of the class A.
Another document, from the same class A as said first document, is picked up (step 310). The region corresponding to the position and size of a potential identifying tag is selected on said other document and a correlation function is computed (step 320) between the potential identifying tag data D and the data of said other document region.
The correlation computation can be expressed for each tag of a model of documents as:
Cr g =1- ( P>>-P'>> ~ ~
With n: number of x coordinates pixels m: number of y coordinates pixels i: incrementation up to value n j: incrementation up to value m P;~: pixel data of the potential identifying tag P';~: pixel data of the corresponding region in the other document The correlation function therefore expresses the similarity between the potential identifying tag and the corresponding region in another document of the same class. If Cog is substantially equal to l, the similarity is high and the potential tag can be considered as being representative of the class A and preselected (step \\HIRSCH6\BREVETSIBrcvets120400\20497EP-dot - 14/03/05 - 15:03 - 6114 330). If not, the potential identifying tag must be rejected (step 340) as not being representative of the class.
The correlation may be refined by computing the function with slight shifts of pixels p' in both coordinate directions x, y and retaining the maximum value.
A
potential identifying tag may be present on another document with a small shift and must be detected anyway.
Such correlation can be conducted for some or each potential tag selected in the first document of the first class and for some or each other document of the same class A.
It will be understood that the routine can be stopped as soon as one match misses. In other words, as soon as one potential identifying tag selected in the first document does not match with one other document of the same class A, said potential identifying tag is rejected (step 340). In one embodiment, the routine can also be stopped as soon as one potential tag is included in one other document of the same class.
If the potential identifying tag matches with at least one other document of class A, said tag can be preselected (step 330). According to embodiments, the check of the potential identifying tag matching with other documents of class A can be conducted on some documents or on all documents of said class A. If class A
only contains one document, all potential tags identified at step 240 will be preselected (step 330).
When an identifying tag is preselected after the series of steps 300, a further series of steps 400 can be conducted to check whether said preselected tag does not match with the documents of other classes, that is, to verify that said preselected tag is discriminant vis-a-vis other classes.
A document of a different class B is picked up (step 410) and a correlation function is computed (step 420) based on the same mathematical expression as set before. If Cog is then substantially equal to 0, the similarity between the preselected identifying tag and a corresponding region of the document of class B is low and said identifying tag can be considered as being discriminant over class B.
Such discrimination control can be conducted over all other classes of documents or stopped as soon as one class B is not represented by said preselected tag.
At least one preselected identifying tags that has not been found in another document of another class is then selected (step 430) as being an identifying tag representing class A otherwise the preselected identifying tag is rejected (step 440).
A model of documents 10 can further be created including said at least one selected identifying tags (i, 2) to represent class A.
\WIRSCH6\BREVETS\Brevets\20400\20497EP,doc - 11/03!05 - 15:03 - 7114 g It must be understood that the steps of selecting potential identifying tags can be mixed with the steps of preselection 300 and selection 400 of an identifying tag. That is to say, when a potential identifying tag is found in one document 8 of one class A, the steps of searching for said identifying tag in other documents of the same S class A and checking that said identifying tag does not match with another document of another class B, can be conducted before another potential identifying tag is found in first document 8.
All the steps of the method for automatically creating a model of documents, as shown in figure 2, can be software implemented.
The invention therefore refers to a computer system for creating a model of documents representing a class of documents comprising computer-implemented means for accessing a plurality of documents separated into different classes;
for determining at least one potential identifying tag within one document belonging to one class; for checking whether said potential identifying tag is included within at least a document of said one class ; for selecting said identifying tag as representing said one class if previous checking is positive; for rejecting said potential identifying tag if previous checking is negative; for creating a model of documents related to said one class including said selected identifying tag.
The computer system may also include computer-implemented means for checking whether said identifying tag is included within at least a document of another class; for selecting said identifying tag as representing said one class if previous checking is negative; for rejecting said preselected identifying tag if previous checking is positive.
The models of documents according to the invention are automatically created with no need for an operator to intervene in the process.
Applicant has run a software according to the method of the invention to create 223 classes starting from an image database of 5 000 digitalized invoices.
Each class includes about 20 documents. This software was run with a Pentium 4 to 2 Giga Hertz and took about 1h30 to create the models.
According to the invention, a set of models of documents is created respectively representing different classes.
Subsequently, when a document to be classified comes in, the models of documents can be used to classify said document, which means to assign a class to said document or to reject said document if not belonging to any known classes. The invention concerns therefore also a method for automatically classifying a document.
Figure 3 is a flow chart illustrating the classification method according to the invention.
\\HIRSCH6\BREVETS\Brevec~\20400\20497EP.doc - 14!03/05 - 1$:03 - &14 There is provided a plurality of models of documents according to the invention (step 500), each model of documents being related to a class.
The document to be classified is compared with a first model of documents refernng to a first class A.
At least one region of the document to be classified is selected, said region corresponding to the size and location of at least one identifying tag of said f rst model of documents. A correlation function is computed (step 510) between said at least one identifying tag and said region of the document to be classified.
The correlation function between the document to be classified and the model of a given class is based on the following expression:
C~ms = ( Crag) ~ Ntag With Crag: the correlation between a tag of the model of documents and a corresponding region of the document to be classified based on the expression set before, and N~g: the number of tags of the model of documents. Nag can be equal or set to 1 according to one embodiment.
If C~~~S is substantially equal to 1, said document matches with said model of documents and can be assigned to the corresponding class (step 520). If C~~~S
is substantially equal to 0, said document does not match with said model of documents.
A subsequent model of documents of another class is picked up (step 530) and the correlation C~nss is calculated (step 540) for said other class.
According to embodiments, the correlation C~~~S can be computed for all defined classes and the document to be classified will be assigned to the class for which C~,~S has the closest value to 1; or the routine will be stopped as soon as one computed value of C~,~S exceeds a threshold value.
If said document to be classified does not match with any model of documents, that is to say if C~tass is smaller than a threshold value substantially close to 0 for all models of documents, said document must be rejected (step 550) as not being suitable for classification according to the method.
All the steps of the method for automatically classifying a document, as shown in figure 3, can be software implemented.
The invention therefore refers to a computer system for classifying a document comprising computer-implemented software means for providing a plurality of models of documents, each model of documents being related to a class; for comparing a document to be classified with one model of documents; for assigning said document to a given class, when said document matches with the corresponding \\HIRSCH6\BREVETS\Brevets\20400\20497EP.doc - 14/03105 - 15:03 - 9l14 one model of documents; for rejecting said document to be classified, if said document to be classified does not match with any model of documents.
The classification method according to the invention increases the rate of automatic document image processing. The document to be classified is solely 5 compared with models of documents by computing data over specific regions of document images corresponding to tags location.
Applicant has run a software according to the method of the invention to classify a document using 223 models created according to the method of the invention. This software was run using a Pentium 4 to 2 Giga Hertz and it took about 10 66,6ms to assign a document to a given class (classification rate is about 15 image documents per second).
1\HIRSCH6\BREVETS\Brevas\20400\20497EP.doc - 14/03/05 - 15.03 - 10/14
Claims (16)
1. A model of documents including at least one identifying tag (1, 2), each identifying tag being defined by at least its graphical content (D), size (S) and location (x, y) within the document.
2. The model of documents according to claim 1, wherein the object to surface ratio of the graphical content of the identifying tag is substantially equal to 0,5.
3. The model of documents according to claim 1 or 2, wherein the size of the identifying tag is less or equal to 5% of the document size.
4. A method for automatically creating a model of documents representing a class of documents, the method comprising the steps of:
a) providing (100) a plurality of documents separated into different classes;
b) determining (200) at least one potential identifying tag within one document (8) belonging to one class (A), said identifying tag (1, 2) being defined by at least its graphical content, size and location within the document;
c) checking whether (300) said potential identifying tag is included within at least a document of said one class (A);
d) if said step c) is positive, selecting (330) said identifying tag;
d') if step c) is negative, rejecting (340) said potential identifying tag and repeating steps b) and c);
e) creating a model of documents (10) related to said one class (A) including said selected identifying tag (1, 2).
a) providing (100) a plurality of documents separated into different classes;
b) determining (200) at least one potential identifying tag within one document (8) belonging to one class (A), said identifying tag (1, 2) being defined by at least its graphical content, size and location within the document;
c) checking whether (300) said potential identifying tag is included within at least a document of said one class (A);
d) if said step c) is positive, selecting (330) said identifying tag;
d') if step c) is negative, rejecting (340) said potential identifying tag and repeating steps b) and c);
e) creating a model of documents (10) related to said one class (A) including said selected identifying tag (1, 2).
5. The method according to claim 4, further comprising the steps of:
f) checking whether (400) said potential identifying tag is included within at least a document of another class (B);
g) if step f) is negative, selecting (430) said identifying tag;
g') if step f) is positive, rejecting (440) said potential identifying tag and repeating steps b) and c).
f) checking whether (400) said potential identifying tag is included within at least a document of another class (B);
g) if step f) is negative, selecting (430) said identifying tag;
g') if step f) is positive, rejecting (440) said potential identifying tag and repeating steps b) and c).
6. Method according to claims 4 or 5, wherein steps b) to g) are repeated for each class of documents provided in step a).
7. Method according to any one of claims 4 to 6, wherein step b) comprises the steps of:
b1) selecting (220) a working window in said one document (8);
b2) calculating (230) an object to surface ratio for said working window;
b3) if said ratio lies within predetermined values, selecting (240) the region of said working window as a potential identifying tag.
b1) selecting (220) a working window in said one document (8);
b2) calculating (230) an object to surface ratio for said working window;
b3) if said ratio lies within predetermined values, selecting (240) the region of said working window as a potential identifying tag.
8. The method according to claim 7, wherein steps b1) to b3) are repeated by shifting the working window to scan (250) the document.
9. The method according to claims 4 or 5, wherein steps c) and f) comprise the step of calculating (320, 420) a correlation function between the potential identifying tag and a region of said document corresponding to the size and location of said identifying tag.
10. The method according to any one of claims 4 to 9, wherein a given portion of the document is excluded from the searching of an identifying tag.
11. A computer system for creating a model of documents representing a class of documents comprising - computer-implemented software means for accessing a plurality of documents separated into different classes;
- computer-implemented software means for determining at least one potential identifying tag within one document belonging to one class, said identifying tag being defined at least by its graphical content, size and location within the document;
- computer-implemented software means for checking whether said potential identifying tag is included within at least another document of said one class;
- computer-implemented software for selecting said identifying tag, if previous checking is positive;
- computer-implemented software for rejecting said potential identifying tag, if previous checking is negative;
- computer-implemented software means for creating a model of documents related to said one class including said selected identifying tag.
- computer-implemented software means for determining at least one potential identifying tag within one document belonging to one class, said identifying tag being defined at least by its graphical content, size and location within the document;
- computer-implemented software means for checking whether said potential identifying tag is included within at least another document of said one class;
- computer-implemented software for selecting said identifying tag, if previous checking is positive;
- computer-implemented software for rejecting said potential identifying tag, if previous checking is negative;
- computer-implemented software means for creating a model of documents related to said one class including said selected identifying tag.
12. The computer system according to claim 11, further comprising:
- computer-implemented software means for checking whether said potential identifying tag is included within at least a document of another class;
- computer-implemented software means for checking whether said potential identifying tag is included within at least a document of another class;
13 - computer-implemented software means for selecting said identifying tag, if previous checking is negative;
- computer-implemented software for rejecting said preselected identifying tag, if previous checking is positive.
13. A method for automatically classifying a document comprising the steps of:
a') providing (500) a plurality of models of documents according to any of claims 4 to 10, each model of documents being related to a given class;
b') comparing (510, 540) a document to be classified with at least one model of documents;
c') if said document to be classified matches with said one model of documents, assigning (520) said document into said corresponding class;
d') if said document to be classified doesn't match with any model of documents, rejecting (550) said document.
- computer-implemented software for rejecting said preselected identifying tag, if previous checking is positive.
13. A method for automatically classifying a document comprising the steps of:
a') providing (500) a plurality of models of documents according to any of claims 4 to 10, each model of documents being related to a given class;
b') comparing (510, 540) a document to be classified with at least one model of documents;
c') if said document to be classified matches with said one model of documents, assigning (520) said document into said corresponding class;
d') if said document to be classified doesn't match with any model of documents, rejecting (550) said document.
14. The method according to claim 13, wherein step b') includes the step of calculating a correlation function between at least one identifying tag of the model of documents and a region of the document to be classified corresponding to the size and location of said identifying tag.
15. The method according to claim 14, wherein step c') includes the steps of:
- comparing the correlation function results for each model of documents provided at step a);
- assigning the document to be classified to the class corresponding to the model of documents providing the correlation result closest to a predetermined value.
- comparing the correlation function results for each model of documents provided at step a);
- assigning the document to be classified to the class corresponding to the model of documents providing the correlation result closest to a predetermined value.
16. A computer system for classifying a document comprising:
- a computer system as set in claim 11 or 12 for providing a plurality of models of documents according to any of claims 1 to 3, each model of documents being related to a given class;
- computer-implemented software means for comparing a document to be classified with at least one model of documents;
- computer-implemented software means for assigning said document to a given class, if said document to be classified matches with the corresponding one model of documents;
- computer-implemented software means for rejecting said document to be classified, if said document to be classified doesn't match with any model of documents.
- a computer system as set in claim 11 or 12 for providing a plurality of models of documents according to any of claims 1 to 3, each model of documents being related to a given class;
- computer-implemented software means for comparing a document to be classified with at least one model of documents;
- computer-implemented software means for assigning said document to a given class, if said document to be classified matches with the corresponding one model of documents;
- computer-implemented software means for rejecting said document to be classified, if said document to be classified doesn't match with any model of documents.
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| EP04290834A EP1583023B1 (en) | 2004-03-30 | 2004-03-30 | Model of documents and method for automatically classifying a document |
| EP04290834.3 | 2004-03-30 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CA2502529A1 true CA2502529A1 (en) | 2005-09-30 |
Family
ID=34878330
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CA002502529A Abandoned CA2502529A1 (en) | 2004-03-30 | 2005-03-29 | Model of documents and method for automatically classifying a document |
Country Status (6)
| Country | Link |
|---|---|
| US (1) | US7370059B2 (en) |
| EP (1) | EP1583023B1 (en) |
| AT (1) | ATE414307T1 (en) |
| AU (1) | AU2005201257A1 (en) |
| CA (1) | CA2502529A1 (en) |
| DE (1) | DE602004017695D1 (en) |
Families Citing this family (18)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20080016052A1 (en) * | 2006-07-14 | 2008-01-17 | Bea Systems, Inc. | Using Connections Between Users and Documents to Rank Documents in an Enterprise Search System |
| US20080016071A1 (en) * | 2006-07-14 | 2008-01-17 | Bea Systems, Inc. | Using Connections Between Users, Tags and Documents to Rank Documents in an Enterprise Search System |
| US20080016061A1 (en) * | 2006-07-14 | 2008-01-17 | Bea Systems, Inc. | Using a Core Data Structure to Calculate Document Ranks |
| US20080016053A1 (en) * | 2006-07-14 | 2008-01-17 | Bea Systems, Inc. | Administration Console to Select Rank Factors |
| US7873641B2 (en) | 2006-07-14 | 2011-01-18 | Bea Systems, Inc. | Using tags in an enterprise search system |
| US20080016072A1 (en) * | 2006-07-14 | 2008-01-17 | Bea Systems, Inc. | Enterprise-Based Tag System |
| JP4910582B2 (en) * | 2006-09-12 | 2012-04-04 | ソニー株式会社 | Information processing apparatus and method, and program |
| JP2008070958A (en) * | 2006-09-12 | 2008-03-27 | Sony Corp | Information processing device and method, and program |
| US8503797B2 (en) * | 2007-09-05 | 2013-08-06 | The Neat Company, Inc. | Automatic document classification using lexical and physical features |
| US8861856B2 (en) | 2007-09-28 | 2014-10-14 | Abbyy Development Llc | Model-based methods of document logical structure recognition in OCR systems |
| US8260049B2 (en) * | 2007-09-28 | 2012-09-04 | Abbyy Software Ltd. | Model-based method of document logical structure recognition in OCR systems |
| US8452132B2 (en) * | 2008-09-23 | 2013-05-28 | Abbyy Software Ltd. | Automatic file name generation in OCR systems |
| US20090292930A1 (en) * | 2008-04-24 | 2009-11-26 | Marano Robert F | System, method and apparatus for assuring authenticity and permissible use of electronic documents |
| US8996350B1 (en) | 2011-11-02 | 2015-03-31 | Dub Software Group, Inc. | System and method for automatic document management |
| US9413912B2 (en) | 2012-10-26 | 2016-08-09 | Abbyy Development Llc | Scanning device having a bed cover including a pattern of repeated design elements |
| JP5572255B1 (en) * | 2013-10-11 | 2014-08-13 | 株式会社Ｕｂｉｃ | Digital information analysis system, digital information analysis method, and digital information analysis program |
| RU2571545C1 (en) * | 2014-09-30 | 2015-12-20 | Общество с ограниченной ответственностью "Аби Девелопмент" | Content-based document image classification |
| US11803918B2 (en) | 2015-07-07 | 2023-10-31 | Oracle International Corporation | System and method for identifying experts on arbitrary topics in an enterprise social network |
Family Cites Families (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US5159667A (en) * | 1989-05-31 | 1992-10-27 | Borrey Roland G | Document identification by characteristics matching |
| US5293429A (en) * | 1991-08-06 | 1994-03-08 | Ricoh Company, Ltd. | System and method for automatically classifying heterogeneous business forms |
| US5917945A (en) * | 1994-06-15 | 1999-06-29 | Metanetics Corporation | Recognizing dataforms in image areas |
| US5937084A (en) * | 1996-05-22 | 1999-08-10 | Ncr Corporation | Knowledge-based document analysis system |
| US6542635B1 (en) * | 1999-09-08 | 2003-04-01 | Lucent Technologies Inc. | Method for document comparison and classification using document image layout |
-
2004
- 2004-03-30 EP EP04290834A patent/EP1583023B1/en not_active Expired - Lifetime
- 2004-03-30 DE DE602004017695T patent/DE602004017695D1/en not_active Expired - Lifetime
- 2004-03-30 AT AT04290834T patent/ATE414307T1/en not_active IP Right Cessation
- 2004-05-25 US US10/852,969 patent/US7370059B2/en active Active
-
2005
- 2005-03-23 AU AU2005201257A patent/AU2005201257A1/en not_active Abandoned
- 2005-03-29 CA CA002502529A patent/CA2502529A1/en not_active Abandoned
Also Published As
| Publication number | Publication date |
|---|---|
| DE602004017695D1 (en) | 2008-12-24 |
| EP1583023B1 (en) | 2008-11-12 |
| ATE414307T1 (en) | 2008-11-15 |
| US20050223313A1 (en) | 2005-10-06 |
| AU2005201257A1 (en) | 2005-10-20 |
| EP1583023A1 (en) | 2005-10-05 |
| US7370059B2 (en) | 2008-05-06 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CA2502529A1 (en) | Model of documents and method for automatically classifying a document | |
| US11004129B2 (en) | Image processing | |
| US8320665B2 (en) | Document image segmentation system | |
| US9965695B1 (en) | Document image binarization method based on content type separation | |
| CN102144236A (en) | Text localization for image and video OCR | |
| JP2010140478A (en) | Method and system for classifying document image | |
| CN110766017A (en) | Mobile terminal character recognition method and system based on deep learning | |
| CN112507782A (en) | Text image recognition method and device | |
| CN104182722A (en) | Text detection method and device and text information extraction method and system | |
| CN110210467B (en) | Formula positioning method of text image, image processing device and storage medium | |
| CN111461131A (en) | Identification method, device, equipment and storage medium for ID card number information | |
| Mishchenko et al. | Model-based chart image classification | |
| US8971669B2 (en) | Method and apparatus for image processing | |
| Badekas et al. | Text binarization in color documents | |
| Lebourgeois et al. | Document analysis in gray level and typography extraction using character pattern redundancies | |
| US20230186592A1 (en) | Division of images into separate color layers | |
| CN114581928A (en) | Form identification method and system | |
| Lang et al. | Physical layout analysis of partly annotated newspaper images | |
| Soua et al. | Improved Hybrid Binarization based on Kmeans for Heterogeneous document processing | |
| Marinai | Learning algorithms for document layout analysis | |
| RU2792722C1 (en) | Splitting images into separate color layers | |
| CN115273061B (en) | Image content level extraction method and system based on principal component analysis | |
| Dhandra et al. | Classification of Document Image Components | |
| Gllavata et al. | Adaptive fuzzy text segmentation in images with complex backgrounds using color and texture | |
| Nikolaou et al. | Text localization in color documents |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| FZDE | Discontinued |