CN107169496A - A kind of character recognition method and device - Google Patents
A kind of character recognition method and device Download PDFInfo
- Publication number
- CN107169496A CN107169496A CN201710257788.8A CN201710257788A CN107169496A CN 107169496 A CN107169496 A CN 107169496A CN 201710257788 A CN201710257788 A CN 201710257788A CN 107169496 A CN107169496 A CN 107169496A
- Authority
- CN
- China
- Prior art keywords
- cutting
- image
- result
- identified
- character
- 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.)
- Pending
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
- G06V30/10—Character recognition
- G06V30/14—Image acquisition
- G06V30/148—Segmentation of character regions
- G06V30/153—Segmentation of character regions using recognition of characters or words
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
- G06F18/2411—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on the proximity to a decision surface, e.g. support vector machines
-
- 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
- G06V30/10—Character recognition
- G06V30/28—Character recognition specially adapted to the type of the alphabet, e.g. Latin alphabet
- G06V30/287—Character recognition specially adapted to the type of the alphabet, e.g. Latin alphabet of Kanji, Hiragana or Katakana characters
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Character Input (AREA)
Abstract
The invention provides a kind of character recognition method and device, methods described includes:Using a variety of slit modes, cutting character image obtains a variety of image cutting results respectively, and multiple cutting regions included by each image cutting result include at least one object to be identified respectively;Object to be identified in the cutting region counts the synthetic weights weight values of each image cutting result, wherein, the synthetic weights weight values include respectively cutting the statistical value of subregional character features matching degree and shape facility matching degree in described image cutting result;Optimum image cutting result is screened according to the synthetic weights weight values of each image cutting result;Recognition result is used as using the corresponding target text in each cutting region in the optimum image cutting result.According to embodiments of the present invention, the recognition accuracy of character image is improved.
Description
Technical field
The present invention relates to field of character recognition, more particularly to a kind of character recognition method, a kind of character recognition device, one
Computer equipment is planted, and, a kind of computer-readable recording medium.
Background technology
In field of character recognition, the especially identification to the character image comprising a whole string literal, it usually needs by whole string
Character segmentation is multiple single words, is identified respectively for each word.
It is typically that cutting, cutting are carried out to image according to the cutting route being made up of multiple cut-offs when cutting
The word merged by multiple words may be included in the region gone out, be identified for the word merged, i.e.,
It can obtain recognition result.
However, in current recognition methods, be present higher error rate in cutting or merging to word, so influence whether
The accuracy of final recognition result.That is, there is the problem of recognition accuracy is relatively low in current Text region mode.
The content of the invention
In view of the above problems, it is proposed that the present invention so as to provide one kind overcome above mentioned problem or at least in part solve on
State a kind of character recognition method of problem, a kind of character recognition device, a kind of computer equipment and a kind of computer-readable deposit
Storage media.
According to one aspect of the present invention there is provided a kind of character recognition method, methods described includes:
Using a variety of slit modes, cutting character image obtains a variety of image cutting results, each image cutting result institute respectively
Including multiple cutting regions respectively include at least one object to be identified;
Object to be identified in the cutting region counts the synthetic weights weight values of each image cutting result, wherein, institute
Synthetic weights weight values are stated including respectively cutting subregional character features matching degree and shape facility matching degree in described image cutting result
Statistical value;
Optimum image cutting result is screened according to the synthetic weights weight values of each image cutting result;
Recognition result is used as using the corresponding target text in each cutting region in the optimum image cutting result.
Alternatively, the object to be identified in the cutting region counts the comprehensive weight of each image cutting result
The step of value, includes:
According to respectively cutting treating in subregional object to be identified and the identification cutting region in described image cutting result
The character features matching degree for the target text that identification object is obtained, and, respectively cut subregional object to be identified and the target
The shape facility matching degree of the corresponding preset shape of word, calculates the synthetic weights weight values of described image cutting result.
Alternatively, it is described using a variety of slit modes cutting character image obtains a variety of image cutting results respectively the step of
Including:
Multiple candidate's cut-offs are marked on the character image;
According to different candidate's cut-offs of selection, multiple candidate's cutting point sets are formed respectively;
According to each candidate's cutting point set, character image described in cutting obtains multiple images cutting result respectively.
Alternatively, the character image includes multiple objects to be slit, described that multiple times are marked on the character image
Cut-off is selected to include following at least one:
Multiple candidate's cut-offs are marked on the impartial multiple positions of character image distance;Or
Each adjacent but disconnected target object to be slit on the character image is searched, and it is to be slit right in each target
Multiple candidate's cut-offs are marked on position as between;Or
Each Object Projection to be slit on the character image is obtained into multiple projection coordinate's points, root in certain direction reference axis
According in certain direction reference axis in the absence of projection coordinate's point coordinate on the character image the corresponding multiple candidates of position mark
Cut-off.
Alternatively, it is described using a variety of slit modes cutting character image obtains a variety of image cutting results respectively the step of
Including:
According to the object to be identified putting in order in the character image, each object to be identified is used successively
Multiple mark windows with different label ranges are marked;
Recognize the target text corresponding to the object to be identified that the mark window of different label ranges is marked;
The object to be identified marked according to the mark window and corresponding target text, filter out each object to be identified
Optimal mark window;
According to the optimal mark window of each object to be identified, character image described in cutting obtains described image cutting knot
Really.
Alternatively, including:
Extract the characteristic vector of the object to be identified in the cutting region;
The word for being matched with the characteristic vector is searched in default characters matching table as the target text;
The cosine value of the object to be identified in the cutting region and the characteristic vector of the target text is calculated, institute is obtained
State and cut subregional character features matching degree.
Alternatively, including:
Search preset shape corresponding with the word classification belonging to the target text;
The cosine value of the ratio of width to height of object to be identified and the preset shape in the cutting region is calculated, obtains described
Cut subregional shape facility matching degree.
Alternatively, subregional object to be identified is respectively cut in the cutting result according to described image with recognizing the cutting
The character features matching degree for the target text that object to be identified in region is obtained, and, respectively cut subregional object to be identified
The shape facility matching degree of preset shape corresponding with the target text, calculates the synthetic weights weight values of described image cutting result
The step of include:
For same image cutting result, calculate multiple average values for cutting subregional character features matching degree and shape is special
The average value of matching degree is levied, respectively as the character features matching degree average and shape facility matching degree of described image cutting result
Average;
By the character features matching degree average and shape facility matching degree average of described image cutting result and the power of distribution
Weight multiplication, and product is summed obtain the synthetic weights weight values of described image cutting result.
According to another aspect of the present invention there is provided a kind of character recognition device, described device includes:
Image cutting result acquisition module, for cutting character image to obtain a variety of images respectively using a variety of slit modes
Cutting result, multiple cutting regions included by each image cutting result include at least one object to be identified respectively;
Comprehensive weight Data-Statistics module, each image cutting knot is counted for the object to be identified in the cutting region
The synthetic weights weight values of fruit;The synthetic weights weight values include respectively cutting subregional character features matching degree in described image cutting result
With the statistical value of shape facility matching degree;
Optimum image cutting result screening module, for screening optimal figure according to the synthetic weights weight values of each image cutting result
As cutting result;
Recognition result determining module, for using the corresponding target text in each cutting region in the optimum image cutting result
Word is used as recognition result.
According to another aspect of the present invention there is provided a kind of computer equipment, the equipment include memory, processor and
Store the computer program that can be run on a memory and on a processor, it is characterised in that journey described in the computing device
The step of any one methods described of above-mentioned character recognition method is realized during sequence.
According to another aspect of the present invention there is provided a kind of computer-readable recording medium, computer journey is stored thereon with
Sequence, it is characterised in that the program realizes the step of any one methods described of above-mentioned character recognition method when being executed by processor
Suddenly.
According to embodiments of the present invention, for being cut using a variety of slit modes a variety of images that cutting character image is obtained respectively
Divide result, the object to be identified in the cutting region counts the synthetic weights weight values of each image cutting result, according to synthesis
The optimum image cutting result that weighted value is filtered out, and using the corresponding target text in each cutting region in optimum image cutting result
Word improves the recognition accuracy of character image as recognition result.
According to embodiments of the present invention, for a variety of image cutting results obtained according to a variety of slit modes, according to word
Characteristic matching degree and shape facility matching degree count the synthetic weights weight values of each image cutting result, are screened according to synthetic weights weight values
The optimum image cutting result gone out.Character features matching degree is introduced during screening and shape facility matching degree is used as ginseng
Examine, both ensured that the word after merging meets target text, deviation is smaller between ensureing the shape of the word after each merging again, makes
The word merging error rate for the image cutting result that must be filtered out is relatively low, so as to improve the recognition accuracy of character image.
According to embodiments of the present invention, it is directed to the multiple images cutting by crossing after cutting after dynamic and being obtained with sliding window identification
As a result, character features matching degree and shape facility matching degree are introduced as the evaluation criterion of image cutting result, by character segmentation
Reasonability quantified by standard of shape facility, and combine character features matching degree, dynamically merge and sliding from crossing after cutting
In the obtained multiple images cutting result of window identification filters out optimum image cutting result, it is to avoid what the mistake to word merged
Overall Text region accuracy rate is in turn ensure that simultaneously, so as to finally improve the accuracy rate of Text region.
Described above is only the general introduction of technical solution of the present invention, in order to better understand the technological means of the present invention,
And can be practiced according to the content of specification, and in order to allow above and other objects of the present invention, feature and advantage can
Become apparent, below especially exemplified by the embodiment of the present invention.
Brief description of the drawings
By reading the detailed description of hereafter preferred embodiment, various other advantages and benefit is common for this area
Technical staff will be clear understanding.Accompanying drawing is only used for showing the purpose of preferred embodiment, and is not considered as to the present invention
Limitation.And in whole accompanying drawing, identical part is denoted by the same reference numerals.In the accompanying drawings:
Fig. 1 is a kind of step flow chart of character recognition method of the embodiment of the present invention one;
Fig. 2 is a kind of step flow chart of character recognition method of the embodiment of the present invention two;
Fig. 3 is a kind of structured flowchart of character recognition device in the embodiment of the present invention three;
Fig. 4 is a kind of structured flowchart of character recognition device in the embodiment of the present invention four;
Fig. 5 is the schematic diagram that a kind of character image of the invention crosses cutting example;
Fig. 6 is the schematic diagram that a kind of character image of the invention dynamically merges example;
Fig. 7 is a kind of schematic flow sheet of sliding window recognition methods of the invention;
Fig. 8 is a kind of slip identification schematic diagram of sliding window identification of the invention;
Fig. 9 is the flow chart that a kind of character recognition method of the invention implements example;
Figure 10 is schematic diagram of the present invention for the alignment score of different images cutting result;
Figure 11 is the schematic diagram one of the ranking results of image cutting result of the present invention;
Figure 12 is the schematic diagram two of the ranking results of image cutting result of the present invention.
Embodiment
The exemplary embodiment of the disclosure is more fully described below with reference to accompanying drawings.Although showing the disclosure in accompanying drawing
Exemplary embodiment, it being understood, however, that may be realized in various forms the disclosure without should be by embodiments set forth here
Limited.On the contrary, these embodiments are provided to facilitate a more thoroughly understanding of the present invention, and can be by the scope of the present disclosure
Complete conveys to those skilled in the art.
Embodiment one
A kind of character recognition method provided in an embodiment of the present invention is discussed in detail.
Reference picture 1, shows a kind of step flow chart of character recognition method in the embodiment of the present invention.
Step 101, using a variety of slit modes, cutting character image obtains a variety of image cutting results respectively, and each image is cut
Multiple cutting regions included by result are divided to include at least one object to be identified respectively.
Above-mentioned character image can be the image comprising text information, comprising text information can for Chinese, English
The word of the multilinguals such as text, French.
Above-mentioned slit mode can include the mode that a variety of words in character image carry out cutting, for example, with
The gap of each word is cut-off in one whole string literal, is cut into multiple words;In another example, will be using word gap as cut-off
Multiple words that cutting is obtained, merging is combined according to neighbouring relations, can obtain different words;In another example, setting is not
With the window of size, character image is scanned according to certain mobile range along words direction, with the motion track of window
Whole string literal is cut into multiple words as cut-off.
Character image is carried out after cutting, character image is cut into multiple cutting including one or more objects to be identified
Subregion, above-mentioned image cutting result is made up of cutting region.Above-mentioned object to be identified can be to have in cutting region
The object of character features, character image is split complete for that behind multiple cutting regions, may include one in each cutting region
Whole word, it is also possible to the part of word, it is also possible to multiple words.
In practical application, continuous cutting region can be understood as the cutting route in character image, according to the cutting road
Footpath carries out cutting, you can to obtain above-mentioned image cutting result.
Same slit mode can obtain multiple images cutting result, according to a variety of slit mode cutting character images,
The multiple images cutting result of a variety of slit modes can be obtained corresponding respectively to.
Step 102, the object to be identified in the cutting region counts the synthetic weights weight values of each image cutting result,
Wherein, the synthetic weights weight values include respectively cutting subregional character features matching degree and shape facility in described image cutting result
The statistical value of matching degree.
In the specific implementation, object to be identified that can be included by each cutting region is identified, will be to be identified right
As being identified as some target text.The concrete mode being identified for cutting subregional object to be identified can have a variety of, example
Such as, each cutting region in character image can be directed to, the characteristic vector of the object to be identified included by it is extracted, by extraction
Characteristic vector is input to SVM (Support Vector Machine, support vector machine) grader, by SVM classifier according to
The characteristic vector of input and the characteristic vector of each default word are compared, and the close word of characteristic vector is output as into target
Word.Certainly, those skilled in the art can according to actual needs adopt and carry out Text region, such as template matches in various manners
Method, geometrical feature extraction method etc..
For recognizing obtained target text, object to be identified included by cutting region and target text can be calculated
Character features matching degree.Character features matching degree can be directed to similarity degree between object to be identified and target text to be counted
Obtained numerical value.Character features matching degree can be obtained in several ways, for example, similar between calculating characteristic vector
Degree, subregional character features matching degree is cut using the similarity as this, character features matching degree show more greatly character features it
Between more match.
Obtained target text is recognized, there can be corresponding preset shape.Different preset shapes has different set
Determine shape facility, such as the depth-width ratio of setting.Shape facility matching degree can be the similarity degree progress for shape between word
Calculate obtained numerical value.For example, the depth-width ratio example of the target text " state " of identification is usually 1.25, and treating in cutting region is known
The depth-width ratio example of other object is 1, and ratio between the two is 1.25, and the difference with 1 for 0.25,1 divided by 0.25 is equal to 4, can be with
Using 4 as shape facility matching degree, shape facility matching degree shows more greatly shape between target text and object to be identified more
Matching.In practical application, corresponding standard shape can be set for each word, can also divided for the word of various language
Not She Ding unified depth-width ratio standard shape, for example, the depth-width ratio example of setting Chinese is 1.2, set the depth-width ratio of Korean as
1.0。
In practical application, especially in the application scenarios of Text region are carried out for identifying code, due to each in identifying code
The shape of individual word is simultaneously nonstandard, can be by the way that the shape of the word in identifying code and preset shape are compared.For example, working as
Some identifying code word is identified as " state ", can be compared the depth-width ratio of identifying code word and the default depth-width ratio of " state "
Compared with, if both depth-width ratio approach, show the identifying code word be " state " probability it is bigger.
Subregional character features matching degree and shape facility matching degree, statistical picture cutting result can be cut according to each
In each cut subregional synthetic weights weight values.The concrete mode of statistics can have a variety of, for example, for the complete of image cutting result
Portion cutting region, calculates each average value for cutting subregional character features matching degree and shape facility matching degree respectively, for
The average value of character features matching degree and the average value of shape facility matching degree, assign weight coefficient, by character features respectively
The product of the average value of average value and shape facility matching degree with degree is summed, and is obtained each in the image cutting result and is cut
Subregional synthetic weights weight values.
Step 103, optimum image cutting result is screened according to the synthetic weights weight values of each image cutting result.
After the synthetic weights weight values of each image cutting result are obtained, it can be filtered out according to the synthetic weights weight values optimal
Image cutting result.The mode of screening can have a variety of, for example, being arranged according to the size of synthetic weights weight values image cutting result
Sequence, forward several image cutting results of sorting as some optimal image cutting results, or, search comprehensive weight
The maximum image cutting result of value is used as optimal image cutting result.
Step 104, tied using the corresponding target text in each cutting region in the optimum image cutting result as identification
Really.
In the specific implementation, the object to be identified that can be directed in each cutting region, which carries out Text region, obtains corresponding target
Word, and using the corresponding target text in each cutting region as the character image recognition result.
According to embodiments of the present invention, for being cut using a variety of slit modes a variety of images that cutting character image is obtained respectively
Divide result, the object to be identified in the cutting region counts the synthetic weights weight values of each image cutting result, according to synthesis
The optimum image cutting result that weighted value is filtered out, and using the corresponding target text in each cutting region in optimum image cutting result
Word improves the recognition accuracy of character image as recognition result.
Embodiment two
Another character recognition method provided in an embodiment of the present invention is discussed in detail.
Reference picture 2, shows the step flow chart of another character recognition method in the embodiment of the present invention.
Step 201, using a variety of slit modes, cutting character image obtains a variety of image cutting results respectively, and each image is cut
Multiple cutting regions included by result are divided to include at least one object to be identified respectively.
Alternatively, the step 201 can include following sub-step:
Sub-step S11, marks multiple candidate's cut-offs on the character image.
Sub-step S12, according to different candidate's cut-offs of selection, forms multiple candidate's cutting point sets respectively.
Sub-step S13, according to each candidate's cutting point set, character image described in cutting obtains multiple images cutting knot respectively
Really.
In the specific implementation, a series of candidate's cut-off can be marked on character image using various ways.More specifically
Ground, can use the mark mode of uniform cutting, for example can be many for mark on the impartial multiple positions of character image distance
Individual candidate's cut-off;Or, the mark mode of connected domain analysis can be used, such as is directed to adjacent but does not connect on character image
Position mark candidate's cut-off between logical image pixel cluster;The mark mode of projective analysis method can also be used, for example
Image pixel on character image is projected in some reference axis, the coordinate for not having projection in reference axis shows in word
In the absence of pixel on the correspondence position of image, the space belonged between word, therefore can be in position mark candidate's cut-off.
Obtain after multiple candidate's cut-offs, can be with selected part candidate cut-off formation candidate's cutting point set.According to choosing
The candidate's cut-off taken is different, then can form multiple different candidate's cutting point sets.
Cutting is carried out to character image according to multiple different candidate's cutting point sets, then can obtain multiple different figures
As cutting result.
In practical application, candidate's cutting point set can be interpreted as a cutting route.From multiple times of mark
Select and one or more candidate's cut-off formation cutting routes are chosen in cut-off.Candidate's cut-off of selection is different, then can be formed
Different cutting routes, carries out cutting to character image according to different cutting routes, can obtain multiple different images and cut
Divide result.
Alternatively, the character image includes multiple objects to be slit, and the sub-step S11 can include following at least one
Kind:
Multiple candidate's cut-offs are marked on the impartial multiple positions of character image distance.
In the specific implementation, can be for marking multiple candidate's cut-offs on the impartial multiple positions of character image distance.Example
Such as, the overall width of character image can be determined first, and overall width divided by the division number of setting are then obtained into target width, with
Target width is the multiple positions difference mark candidate cut-off being spaced on character image.
Each adjacent but disconnected target object to be slit on the character image is searched, and it is to be slit right in each target
Multiple candidate's cut-offs are marked on position as between.
In the specific implementation, can include in character image by different image pixel clusters constitute it is multiple to be slit right
As.Carry out connected domain analysis for each object to be slit, determine whether connected between adjacent object to be slit, i.e., adjacent two
With the presence or absence of the image pixel being connected with each other between individual image pixel cluster.It regard adjacent but disconnected object to be slit as mesh
Mark object to be slit, and mark candidate cut-off on the position between target object to be slit.
Each Object Projection to be slit on the character image is obtained into multiple projection coordinate's points, root in certain direction reference axis
According in certain direction reference axis in the absence of projection coordinate's point coordinate on the character image the corresponding multiple candidates of position mark
Cut-off.
In the specific implementation, can be directed to character arranging direction in character image sets a reference axis laterally or longitudinally,
By the multiple images pixel projection in word graph image in reference axis, so as to obtain one in multiple coordinate points of reference axis and be
The subpoint corresponding to each image pixel of row, if subpoint is not present in some coordinate points, shows the correspondence in character image
In the absence of image pixel on position, the space belonged between word, therefore can be in position mark candidate's cut-off.
It should be noted that in actual applications, can using it is above-mentioned it is one or more by the way of mark candidate cutting
Point.
In actual applications, the method dynamically merged after cutting can be used to realize that above-mentioned steps obtain multiple images and cut
Divide result.Specifically, cutting can be carried out to object to be identified first.For object to be identified in character image it is residing
Position, by uniform cutting, projective analysis method and connected domain analysis method mark as far as possible by the candidate of object to be identified cutting
Cut-off, cutting result was obtained by multiple candidate's cut-offs.Cutting granularity can be controlled in dicing process, will be treated with realizing
Identification object cutting as much as possible.Fig. 5 shows that a kind of character image crosses the schematic diagram of cutting example.It can be seen that being directed to
Input picture " wide along North Street ", each do not connect or between exist space object to be identified or object to be identified side by marked
Remember and candidate's cut-off, so as to obtain cutting result.
Cutting result is crossed according to what is obtained, multiple images cutting result can be obtained by dynamic merging.Select several
Candidate's cut-off, cutting is carried out on the basis of candidate's cut-off of selection to object to be identified, generate by several words or
The object to be identified merged by word side.The different candidate's cut-off of selection, then can produce different word amalgamation results.
Multiple images cutting result is obtained by different word amalgamation results.Fig. 6 shows that a kind of character image dynamically merges example
Schematic diagram.It can be seen that being directed to cutting result, selected candidate's cut-off is different, then can produce different dynamic conjunctions
And result, that is, obtain multiple image cutting results.
Alternatively, the step 201 can include following sub-step:
Sub-step S21, according to the object to be identified putting in order in the character image, waits to know successively to each
Other object is marked using multiple mark windows with different label ranges.
Sub-step S22, recognizes the target text corresponding to the object to be identified that the mark window of different label ranges is marked.
Sub-step S23, according to character features of the object to be identified of mark window mark with corresponding target text
Matching degree, filters out the optimal mark window of each object to be identified.
Sub-step S24, according to the optimal mark window of each object to be identified, character image described in cutting obtains the figure
As cutting result.
The above-mentioned mark window with label range can be the window with certain altitude and width, for by window
Included one or more objects to be identified are marked.In the specific implementation, multiple different label ranges can be pre-set
Mark window, according to putting in order for object to be identified, each object to be identified is marked successively.Different label ranges
Mark window included by object to be identified it is different, i.e., the object to be identified that the mark windows of different label ranges is marked is not
Together.For example, for object to be identified " suitable ", the mark window of a large-size can enter rower for whole word " suitable "
Note, and then object to be identified " river " and object to be identified " page " can be marked respectively for the mark window of reduced size.
It is identified for the object to be identified marked in mark window, obtains corresponding target text.For mark
The method that object to be identified is identified has a variety of, for example, the characteristic vector of the object to be identified included by mark window is extracted,
The characteristic vector of extraction is input to SVM classifier, by characteristic vector and each default word of the SVM classifier according to input
Characteristic vector is compared, and the close word of characteristic vector is output as into target text.
The character features matching degree between the object to be identified of mark and the target text of identification can be calculated.Character features
The specific computational methods of matching degree can have a variety of.For example, calculating between target text and the characteristic vector of object to be identified
Similarity, using the similarity as character features matching degree, character features matching degree shows more to match between character features more greatly.
If the character features matching degree between object to be identified and target text that current mark window is marked is more than
Predetermined threshold value, or sort in the mark window of multiple different label ranges it is forward, can be as optimal mark window.
It is thus possible to from multiple mark windows with different label ranges, filter out one or more optimal mark windows.Press
Label range according to optimal mark window carries out cutting to character image, obtains multiple images cutting result.
In actual applications, method for distinguishing can be known using sliding window and realizes that above-mentioned steps obtain multiple images cutting result.
Fig. 7 shows a kind of schematic flow sheet of sliding window recognition methods.It can be seen that for the character picture as character image,
Putting in order for object to be identified can be determined first, so as to the glide direction put in order according to this as window.It is actual to answer
, can be according to common ways of writing default window glide direction for from left to right in.It can be determined by Gray Projection method
Left margin in character image residing for object to be identified, so as to which the left margin of object to be identified is started as current location
Sliding window.
The reference dimension of window can be determined by analyzing the width distribution of overall object to be identified., can in practical application
To set a word training sample set, the distribution [13,35] for the text width concentrated according to word training sample can be with
The width of the multiple windows of respective settings is respectively 12,14,16,18,20,22,24,26,28,30,32,34,36 totally 13 grades.For
The window of different in width, enters line slip according to certain slip amplitude along direction from left to right, in sliding process, for
Image in current window carries out Text region.For example, the image in current window region is input into the good volume of training in advance
Product neutral net, the characteristic vector value of extraction is output as by it.The characteristic vector value of extraction is inputted to SVM classifier, obtained
Classification results.
The characteristic vector value that the object to be identified that can be marked with calculation window is extracted, with differentiating generic corresponding one
Cosine value between the characteristic vector value of serial prototype, obtains character features matching degree.
After current object to be identified is identified using the window of multiple different in width, each window is obtained corresponding
Character features matching degree.Enter line slip according to the maximum window of character features matching degree, until sliding into the right side of object to be identified
Border, then slide and terminate and output character recognition result.Fig. 8 shows a kind of slip identification schematic diagram of sliding window identification.From figure
In it is visible,, can be to different in width when the searching currently optimal window by starting point of " wide " word for the word of " wide along North Street "
Window in picture material be identified, according to character features matching degree obtain mark word " suitable " window be optimal window
Mouthful.
In practical application, parameter and SVM classifier in convolutional neural networks can train sample according to the word of mark
This collection is optimized and machine training.Wherein, the characteristic vector value of generic correspondence prototype can be in machine learning process
Generation.For example, convolutional neural networks can carry out machine training using substantial amounts of word training sample.Wherein convolutional neural networks
Improved Le-Net structures can be used, the structure includes four convolutional layers, four pond layers and two full articulamentums, last
Layer is output as the characteristic vector of 512 dimensions.Chinese Character Recognition is directed to, the class object of SVM classifier can be set as 3755 classes one
Level Chinese character, and the word training sample for being used for the mark of each class of machine training is 200, the precision of word training sample set can
Think 99.2%.A series of characteristic vector value of prototypes corresponding for generic, can first by word training sample according to
A series of secondary characteristic vector value that 512 dimensions are generated by the convolutional neural networks trained, then to the characteristic vector value of generation
Clustered, cluster the classification prototype that obtained class center is each class.For K class problems, control convergence can be passed through
Parameter, each class Ci finally given (i=1,2 ... K) prototype number Ni can be different.Layer can be selected in actual cluster
Secondary clustering method.
It should be noted that the character recognition method of the embodiment of the present invention can be applied to a variety of differences according to actual needs
Slit mode, however it is not limited to above-mentioned cross dynamically merges and slit mode that sliding window is recognized after cutting.
Step 202, according to respectively cutting subregional object to be identified in described image cutting result with recognizing the cutting area
The character features matching degree for the target text that object to be identified in domain is obtained, and, respectively cut subregional object to be identified with
The shape facility matching degree of the corresponding preset shape of the target text, calculates the synthetic weights weight values of described image cutting result.
Alternatively, the step 202 can include following sub-step:
Sub-step S31, extracts the characteristic vector of the object to be identified in the cutting region.
Sub-step S32, searches the word for being matched with the characteristic vector as the mesh in default characters matching table
Mark word.
Sub-step S33, calculates the remaining of object to be identified in the cutting region and the characteristic vector of the target text
String value, obtains described cutting subregional character features matching degree.
In the specific implementation, object to be identified that can be included by each cutting region carries out the extraction of characteristic vector.
Characteristic vector can be by position coordinate data of the object to be identified in character image constituted be used for express character features
A series of vectors.For the characteristic vector of extraction, the matching target text of characteristic vector can be searched.For cutting area
The characteristic vector of the characteristic vector of object to be identified included by domain and the target text found calculates cosine value, by what is obtained
The cosine value cutting region is used as character features matching degree.
Alternatively, the step 202 can also include following sub-step:
Sub-step S34, searches preset shape corresponding with the word classification belonging to the target text.
Object to be identified and the cosine of the ratio of width to height of the preset shape in sub-step S35, the calculating cutting region
Value, obtains described cutting subregional shape facility matching degree.
In the specific implementation, the preset shape of its affiliated word classification can be set for target text.Preset shape can be with
Word the ratio of width to height or other attribute informations for being used to represent word shape for standard.Chinese character can be for example directed to
It is 0.8 to set the other the ratio of width to height of Chinese characters kind.For the ratio of width to height and corresponding preset shape of the word included by cutting region
The ratio of width to height calculates cosine value, regard obtained cosine value as shape facility matching degree.
Above-mentioned character features matching degree and shape facility matching degree closer to 1, show character features between word and
Shape facility is more matched.
Alternatively, the step 202 can include following sub-step:
Sub-step S41, for same image cutting result, calculating is multiple to cut being averaged for subregional character features matching degree
The average value of value and shape facility matching degree, respectively as the character features matching degree average and shape of described image cutting result
Characteristic matching degree average.
Sub-step S42, by the character features matching degree average of described image cutting result and shape facility matching degree average
It is multiplied with the weight coefficient of distribution, and product is summed obtains the synthetic weights weight values of described image cutting result.
In the specific implementation, for whole cutting region of image cutting result, each being calculated respectively and cuts subregional word
The average value of characteristic matching degree and shape facility matching degree.
The average value of average value and shape facility matching degree for character features matching degree, is multiplied by the weight of setting respectively
Coefficient, the product of the average value and weight coefficient of the average value of character features matching degree and shape facility matching degree is asked
With obtain each in the image cutting result and cut subregional synthetic weights weight values.
In practical application, weight coefficient can be obtained by way of the adjusting parameter on word training sample set.It is optional
Ground, character features matching degree and shape facility matching degree can assign weight coefficient 0.8 and 0.2 respectively.
In actual applications, using some or multiple candidate's cutting point set cutting character images, obtain some or
, can be by the image cutting result newly obtained with being entered before this using other candidate's cutting point sets after multiple images cutting result
Other image cutting results that row cutting is obtained, are arranged from big to small according to character features matching degree or shape facility matching degree
Sequence, only retains M image cutting result before sequence.The renewal being ranked up for image cutting result newly-increased every time, directly
To all candidate's cutting point sets are traveled through, M image cutting result is finally given.
Generally, crossing cutting dynamically to merge is gone forward side by side based on Beam Search (beam-search) search selection candidate's cut-offs
Mobile state merges, if there is N number of candidate's cut-off, can produce 2NIndividual image cutting result, therefore can use above-mentioned
The mode of M image cutting result realizes Pruning strategy before sorting and retaining.So as to screen optimal figure according to synthetic weights weight values
During as cutting result, screened without to substantial amounts of image cutting result, improve the speed of Text region.
In actual applications, can also be by the cosine value between the ratio of width to height for respectively cutting subregional object to be identified with presetting
Threshold value is compared, and when cosine value is less than predetermined threshold value, can be filtered out corresponding image cutting result.When same image is cut
Cosine value that each in point result is cut between subregional the ratio of width to height is less than predetermined threshold value, shows right in present image cutting result
Be present mistake in the cutting and merging of object to be identified, therefore can be filtered out.In practical application, it can be existed by above-mentioned steps
Increase merges limitation when producing image cutting result, it is to avoid the word for producing mistake merges.
By the way that the undesirable image cutting result of the ratio of width to height is filtered out, it is to avoid invalid image cutting result is entered
The follow-up sequence of row, improves the speed of Text region.
Step 203, optimum image cutting result is screened according to the synthetic weights weight values of each image cutting result.
In the specific implementation, can be from big to small ranked up according to the synthetic weights weight values of image cutting result, by synthetic weights
The maximum image cutting result of weight values is used as optimum image cutting result.
Step 204, tied using the corresponding target text in each cutting region in the optimum image cutting result as identification
Really.
It should be noted that crossing in the identification method dynamically merged after cutting, for target text region, first pass through certain
Rule determine as far as possible by the cut-off of object cutting, a cutting road is corresponded to by any one subset of cutting point set
Footpath, carries out cutting to the object in character image according to cutting route, several object mergings in cut-off is treated into one
Identification object is simultaneously identified.However, dynamic when merging may mistakenly combining objects, so as to cause Text region mistake,
It has impact on the accuracy rate of Text region.
In the identification method of sliding window identification, the window of sizes is typically set, and along specific direction to word graph
As being scanned, the character image scanned to specific dimensions window carries out Text region.Know when using the window of certain size
Do not go out after word, then cutting route is formed using the border of the window of the size as cut-off, follow-up object to be identified is entered
Row cutting is simultaneously recognized.However, sliding window identification carries out cutting and merging only in accordance with the window of local optimum to whole objects, lack
Global information is easily trapped into local optimum in the case of instructing, it is impossible to ensures the accuracy rate of other parts Text region, have impact on
The accuracy rate of Text region.
According to embodiments of the present invention, for a variety of image cutting results obtained according to a variety of slit modes, according to word
Characteristic matching degree and shape facility matching degree count the synthetic weights weight values of each image cutting result, are screened according to synthetic weights weight values
The optimum image cutting result gone out.Character features matching degree is introduced during screening and shape facility matching degree is used as ginseng
Examine, both ensured that the word after merging meets target text, deviation is smaller between ensureing the shape of the word after each merging again, makes
The word merging error rate for the image cutting result that must be filtered out is relatively low, so as to improve the recognition accuracy of character image.
According to embodiments of the present invention, it is directed to the multiple images cutting by crossing after cutting after dynamic and being obtained with sliding window identification
As a result, character features matching degree and shape facility matching degree are introduced as the evaluation criterion of image cutting result, by character segmentation
Reasonability quantified by standard of shape facility, and combine character features matching degree, dynamically merge and sliding from crossing after cutting
In the obtained multiple images cutting result of window identification filters out optimum image cutting result, it is to avoid what the mistake to word merged
Overall Text region accuracy rate is in turn ensure that simultaneously, so as to finally improve the accuracy rate of Text region.
Deeply understand the embodiment of the present invention for the ease of those skilled in the art, carried out below with reference to specific implementation example
Explanation.
Fig. 9 shows that a kind of character recognition method of the invention implements the flow chart of example.It can be seen that for input
Image, can be pre-processed first.Pretreatment can include gray processing, noise reduction, binaryzation, character cutting and normalization
These process steps.After binaryzation, image is only left two kinds of colors, i.e., black and white, one of them is image background, another
Individual color seeks to the word of identification.Character cutting is then into single word by the Text segmentation in image.If word is tilted,
Line tilt correction can be entered.Normalization is then to arrive same size by single character image is regular.
Pretreated image, can be respectively according to dynamically merging and sliding window knowledge obtain multiple figures otherwise after cutting excessively
As cutting result.According to character features matching degree and the synthetic weights weight values of shape facility matching degree statistical picture cutting result, root
Paths ordering is carried out to multiple images cutting result according to synthetic weights weight values, will be sorted corresponding to forward optimum image cutting result
Target text be used as recognition result.
Figure 10 shows schematic diagram of the present invention for the alignment score of different images cutting result.It can be seen that pin
To the character features matching degree and shape facility matching degree of multiple images cutting result, after being multiplied respectively with the weight coefficient of setting
It is added, obtains synthetic weights weight values.The image cutting result " wide along North Street " that synthetic weights weight values are 0.84 is cut for optimal image
Divide result, and in the image cutting result of " wide river page North Street ", although its character features matching degree is 0.85, but its cutting
Word shape is unreasonable, and shape facility matching degree only has 0.72 point, and its final synthetic weights weight values is 0.82.
Figure 11 shows the schematic diagram one of the ranking results of image cutting result of the present invention.It can be seen that for word
Image " Beijing-Xicheng District De Wai street religion ", used the image cutting result " Beijing-west city for dynamically merging after cutting and obtaining
The outer street religions of Qu De ", its synthetic weights weight values is 0.86.And obtained image cutting result " Beijing-Xicheng District is recognized by sliding window
Teach in De Xibu street ", due to combinde rqdical character " outer " to have been carried out to the cutting of mistake, its synthetic weights weight values is 0.83.Therefore can be by
" Beijing-Xicheng District De Wai street religion " is used as optimum image cutting result.
Figure 12 shows the schematic diagram two of the ranking results of image cutting result of the present invention.It can be seen that for word
Image " icehouse (river is big ", the image cutting result " icehouse Dil is big " for dynamically merging after cutting and obtaining was used, due to its cutting
The word shape of merging is unreasonable, and its synthetic weights weight values is 0.72.And obtained image cutting result " icehouse is recognized by sliding window
(river is big " synthetic weights weight values be 0.88, can be as optimum image cutting result.
Pass through above-mentioned specific example, it is seen that the embodiment of the present invention can be realized dynamically to be merged and sliding window knowledge to crossing after cutting
The complementation of other two schemes, had not only ensured local optimum but also can guarantee that global optimum, while also improving recognition efficiency.In actual reality
In testing, in the test sample comprising 2000 Chinese block letter images, crossing the recognition accuracy dynamically merged after cutting is
70%, the recognition accuracy of sliding window identification is 76%, and uses the Text region mode of the embodiment of the present invention, its recognition accuracy
For 83%.
It should be noted that for foregoing embodiment of the method, in order to be briefly described, therefore it is all expressed as a series of
Combination of actions, but those skilled in the art should know, the present invention is not limited by described sequence of movement, because according to
According to the present invention, some steps can be carried out sequentially or simultaneously using other.Secondly, those skilled in the art should also know,
Embodiment described in this description belongs to preferred embodiment, and involved action is not necessarily essential to the invention.
Embodiment three
A kind of character recognition device provided in an embodiment of the present invention is discussed in detail.
Reference picture 3, shows a kind of structured flowchart of character recognition device in the embodiment of the present invention three.
Described device can include:
Image cutting result acquisition module 301, for cutting character image to obtain a variety of respectively using a variety of slit modes
Image cutting result, multiple cutting regions included by each image cutting result include at least one object to be identified respectively;
Comprehensive weight Data-Statistics module 302, counts each image for the object to be identified in the cutting region and cuts
Divide the synthetic weights weight values of result, wherein, the synthetic weights weight values include respectively cutting subregional word in described image cutting result
The statistical value of characteristic matching degree and shape facility matching degree;
Optimum image cutting result screening module 303, for being screened most according to the synthetic weights weight values of each image cutting result
Excellent image cutting result;
Recognition result determining module 304, for using the corresponding mesh in each cutting region in the optimum image cutting result
Mark word is used as recognition result.
According to embodiments of the present invention, for being cut using a variety of slit modes a variety of images that cutting character image is obtained respectively
Divide result, the object to be identified in the cutting region counts the synthetic weights weight values of each image cutting result, according to synthesis
The optimum image cutting result that weighted value is filtered out, and using the corresponding target text in each cutting region in optimum image cutting result
Word improves the recognition accuracy of character image as recognition result.
Example IV
A kind of character recognition device provided in an embodiment of the present invention is discussed in detail.
Reference picture 4, shows a kind of structured flowchart of character recognition device in the embodiment of the present invention four.
Described device can include:
Image cutting result acquisition module 401, for cutting character image to obtain a variety of respectively using a variety of slit modes
Image cutting result, multiple cutting regions included by each image cutting result include at least one object to be identified respectively.
Synthetic weights weight values computing module 402, for subregional to be identified right according to respectively being cut in described image cutting result
As the character features matching degree with recognizing the target text that the object to be identified in the cutting region is obtained, and, each cutting
The shape facility matching degree of the object to be identified in region preset shape corresponding with the target text, calculates described image cutting
As a result synthetic weights weight values.
Image cutting result screening module 403, for screening optimal figure according to the synthetic weights weight values of each image cutting result
As cutting result.
Recognition result determining module 404, for using the corresponding mesh in each cutting region in the optimum image cutting result
Mark word is used as recognition result.
Alternatively, described image cutting result acquisition module 401 includes:
Candidate's cut-off marks submodule, for marking multiple candidate's cut-offs on the character image;
Candidate's cutting point set formation submodule, for different candidate's cut-offs according to selection, forms multiple times respectively
Select cutting point set;
Cutting character image submodule, for according to each candidate's cutting point set, character image described in cutting to be obtained respectively
Multiple images cutting result.
Alternatively, the character image includes multiple objects to be slit, and candidate's cut-off mark submodule is included such as
Lower at least one:
First candidate's cut-off marks subelement, many for being marked on the impartial multiple positions of character image distance
Individual candidate's cut-off;
Second candidate's cut-off marks subelement, for searching each adjacent but disconnected target on the character image
Multiple candidate's cut-offs are marked on object to be slit, and position between each target object to be slit;
3rd candidate's cut-off mark subelement, for by each Object Projection to be slit on the character image in certain direction
Multiple projection coordinate's points are obtained in reference axis, according in certain direction reference axis in the absence of projection coordinate's point coordinate in the word
The corresponding multiple candidate's cut-offs of position mark on image.
Alternatively, described image cutting result acquisition module 401 includes:
Window indicia submodule, for the putting in order in the character image according to the object to be identified, successively
Each object to be identified is marked using multiple mark windows with different label ranges;
Target text recognizes submodule, for recognizing that the object to be identified institute that the mark window of different label ranges is marked is right
The target text answered;
Optimal mark window screens submodule, for object to be identified and the corresponding mesh marked according to the mark window
Word is marked, the optimal mark window of each object to be identified is filtered out;
Mark window cutting character image submodule, for the optimal mark window according to each object to be identified, cutting
The character image obtains described image cutting result.
Alternatively, the synthetic weights weight values computing module 402 includes:
Characteristic vector pickup submodule, the characteristic vector for extracting the object to be identified in the cutting region;
Target text searches submodule, and the text of the characteristic vector is matched with for being searched in default characters matching table
Word is used as the target text;
First cosine value calculating sub module, for calculating the object to be identified in the cutting region and the target text
Characteristic vector cosine value, obtain described cutting subregional character features matching degree.
Alternatively, the synthetic weights weight values computing module 402 includes:
Preset shape searches submodule, for searching default shape corresponding with the word classification belonging to the target text
Shape;
Second cosine value calculating sub module, for calculating the object to be identified in the cutting region and the preset shape
The ratio of width to height cosine value, obtain described cutting subregional shape facility matching degree.
Alternatively, the synthetic weights weight values computing module 402 includes:
Mean value calculation submodule, for for same image cutting result, calculating is multiple to cut subregional character features
The average value of matching degree and the average value of shape facility matching degree, are matched respectively as the character features of described image cutting result
Spend average and shape facility matching degree average;
Product is summed submodule, for by the character features matching degree average and shape facility of described image cutting result
It is multiplied with degree average with the weight coefficient of distribution, and product summation is obtained into the synthetic weights weight values of described image cutting result.
According to embodiments of the present invention, for a variety of image cutting results obtained according to a variety of slit modes, according to word
Characteristic matching degree and shape facility matching degree count the synthetic weights weight values of each image cutting result, are screened according to synthetic weights weight values
The optimum image cutting result gone out.Character features matching degree is introduced during screening and shape facility matching degree is used as ginseng
Examine, both ensured that the word after merging meets target text, deviation is smaller between ensureing the shape of the word after each merging again, makes
The word merging error rate for the image cutting result that must be filtered out is relatively low, so as to improve the recognition accuracy of character image.
According to embodiments of the present invention, it is directed to the multiple images cutting by crossing after cutting after dynamic and being obtained with sliding window identification
As a result, character features matching degree and shape facility matching degree are introduced as the evaluation criterion of image cutting result, by character segmentation
Reasonability quantified by standard of shape facility, and combine character features matching degree, dynamically merge and sliding from crossing after cutting
In the obtained multiple images cutting result of window identification filters out optimum image cutting result, it is to avoid what the mistake to word merged
Overall Text region accuracy rate is in turn ensure that simultaneously, so as to finally improve the accuracy rate of Text region.
For above-mentioned character recognition device embodiment, because it is substantially similar to embodiment of the method, so description
Fairly simple, the relevent part can refer to the partial explaination of embodiments of method.
Embodiment five
A kind of computer equipment provided in an embodiment of the present invention and a kind of computer-readable recording medium is discussed in detail.
The computer equipment includes memory, processor and stores the meter that can be run on a memory and on a processor
Calculation machine program, it is characterised in that following steps can be realized during the computing device described program:
Using a variety of slit modes, cutting character image obtains a variety of image cutting results, each image cutting result institute respectively
Including multiple cutting regions respectively include at least one object to be identified;Object to be identified system in the cutting region
The synthetic weights weight values of each image cutting result are counted, wherein, the synthetic weights weight values include each cutting in described image cutting result
The character features matching degree in region and the statistical value of shape facility matching degree;Sieved according to the synthetic weights weight values of each image cutting result
Select optimum image cutting result;Identification is used as using the corresponding target text in each cutting region in the optimum image cutting result
As a result.
Alternatively, following steps can also be realized during the computing device described program:
According to respectively cutting treating in subregional object to be identified and the identification cutting region in described image cutting result
The character features matching degree for the target text that identification object is obtained, and, respectively cut subregional object to be identified and the target
The shape facility matching degree of the corresponding preset shape of word, calculates the synthetic weights weight values of described image cutting result.
Alternatively, following steps can also be realized during the computing device described program:
Multiple candidate's cut-offs are marked on the character image;According to different candidate's cut-offs of selection, formed respectively
Multiple candidate's cutting point sets;According to each candidate's cutting point set, character image described in cutting obtains multiple images cutting respectively
As a result.
Alternatively, following steps can also be realized during the computing device described program:
The character image includes multiple objects to be slit, described that multiple candidate's cut-offs are marked on the character image
Including following at least one:Multiple candidate's cut-offs are marked on the impartial multiple positions of character image distance;Or, look into
Look for each adjacent but disconnected target object to be slit on the character image, and the position between each target object to be slit
Put the multiple candidate's cut-offs of mark;Or, by each Object Projection to be slit on the character image in certain direction reference axis
Obtain multiple projection coordinate's points, according in certain direction reference axis in the absence of projection coordinate's point coordinate on the character image it is right
The multiple candidate's cut-offs of position mark answered.
Alternatively, following steps can also be realized during the computing device described program:
According to the object to be identified putting in order in the character image, each object to be identified is used successively
Multiple mark windows with different label ranges are marked;Recognize different label ranges mark window mark it is to be identified
Target text corresponding to object;The object to be identified marked according to the mark window and corresponding target text, are filtered out
The optimal mark window of each object to be identified;According to the optimal mark window of each object to be identified, word graph described in cutting
As obtaining described image cutting result.
Alternatively, following steps can also be realized during the computing device described program:
Extract the characteristic vector of the object to be identified in the cutting region;Matching is searched in default characters matching table
The target text is used as in the word of the characteristic vector;Calculate the object to be identified and the target in the cutting region
The cosine value of the characteristic vector of word, obtains described cutting subregional character features matching degree.
Alternatively, following steps can also be realized during the computing device described program:
Search preset shape corresponding with the word classification belonging to the target text;Calculate treating in the cutting region
The cosine value of the ratio of width to height of identification object and the preset shape, obtains described cutting subregional shape facility matching degree.
Alternatively, following steps can also be realized during the computing device described program:
For same image cutting result, calculate multiple average values for cutting subregional character features matching degree and shape is special
The average value of matching degree is levied, respectively as the character features matching degree average and shape facility matching degree of described image cutting result
Average;By the character features matching degree average and shape facility matching degree average of described image cutting result and the weight system of distribution
Number is multiplied, and product is summed obtains the synthetic weights weight values of described image cutting result.
The computer-readable recording medium storage has computer program, can be realized such as when the program is executed by processor
Lower step:
Using a variety of slit modes, cutting character image obtains a variety of image cutting results, each image cutting result institute respectively
Including multiple cutting regions respectively include at least one object to be identified;Object to be identified system in the cutting region
The synthetic weights weight values of each image cutting result are counted, wherein, the synthetic weights weight values include each cutting in described image cutting result
The character features matching degree in region and the statistical value of shape facility matching degree;Sieved according to the synthetic weights weight values of each image cutting result
Select optimum image cutting result;Identification is used as using the corresponding target text in each cutting region in the optimum image cutting result
As a result.
Alternatively, following steps can also be realized when the program is executed by processor:
According to respectively cutting treating in subregional object to be identified and the identification cutting region in described image cutting result
The character features matching degree for the target text that identification object is obtained, and, respectively cut subregional object to be identified and the target
The shape facility matching degree of the corresponding preset shape of word, calculates the synthetic weights weight values of described image cutting result.
Alternatively, following steps can also be realized when the program is executed by processor:
Multiple candidate's cut-offs are marked on the character image;According to different candidate's cut-offs of selection, formed respectively
Multiple candidate's cutting point sets;According to each candidate's cutting point set, character image described in cutting obtains multiple images cutting respectively
As a result.
Alternatively, following steps can also be realized when the program is executed by processor:
The character image includes multiple objects to be slit, described that multiple candidate's cut-offs are marked on the character image
Including following at least one:Multiple candidate's cut-offs are marked on the impartial multiple positions of character image distance;Or, look into
Look for each adjacent but disconnected target object to be slit on the character image, and the position between each target object to be slit
Put the multiple candidate's cut-offs of mark;Or, by each Object Projection to be slit on the character image in certain direction reference axis
Obtain multiple projection coordinate's points, according in certain direction reference axis in the absence of projection coordinate's point coordinate on the character image it is right
The multiple candidate's cut-offs of position mark answered.
Alternatively, following steps can also be realized when the program is executed by processor:
According to the object to be identified putting in order in the character image, each object to be identified is used successively
Multiple mark windows with different label ranges are marked;Recognize different label ranges mark window mark it is to be identified
Target text corresponding to object;The object to be identified marked according to the mark window and corresponding target text, are filtered out
The optimal mark window of each object to be identified;According to the optimal mark window of each object to be identified, word graph described in cutting
As obtaining described image cutting result.
Alternatively, following steps can also be realized when the program is executed by processor:
Extract the characteristic vector of the object to be identified in the cutting region;Matching is searched in default characters matching table
The target text is used as in the word of the characteristic vector;Calculate the object to be identified and the target in the cutting region
The cosine value of the characteristic vector of word, obtains described cutting subregional character features matching degree.
Alternatively, following steps can also be realized when the program is executed by processor:
Search preset shape corresponding with the word classification belonging to the target text;Calculate treating in the cutting region
The cosine value of the ratio of width to height of identification object and the preset shape, obtains described cutting subregional shape facility matching degree.
Alternatively, following steps can also be realized when the program is executed by processor:
For same image cutting result, calculate multiple average values for cutting subregional character features matching degree and shape is special
The average value of matching degree is levied, respectively as the character features matching degree average and shape facility matching degree of described image cutting result
Average;By the character features matching degree average and shape facility matching degree average of described image cutting result and the weight system of distribution
Number is multiplied, and product is summed obtains the synthetic weights weight values of described image cutting result.
Each embodiment in this specification is described by the way of progressive, what each embodiment was stressed be with
Between the difference of other embodiment, each embodiment identical similar part mutually referring to.
It would have readily occurred to a person skilled in the art that be:Any combination application of each above-mentioned embodiment is all feasible, therefore
Any combination between each above-mentioned embodiment is all embodiment of the present invention, but this specification exists as space is limited,
This is not just detailed one by one.
Provided herein based on mobile terminal phone report scheme not with any certain computer, virtual system or its
Its equipment is inherently related.Various general-purpose systems can also be used together with based on teaching in this.As described above, construct
It is obvious with the structure required by the system of the present invention program.In addition, the present invention is not also directed to any certain programmed
Language.It is understood that, it is possible to use various programming languages realize the content of invention described herein, and above to specific language
The done description of speech is to disclose the preferred forms of the present invention.
In the specification that this place is provided, numerous specific details are set forth.It is to be appreciated, however, that the implementation of the present invention
Example can be put into practice in the case of these no details.In some instances, known method, structure is not been shown in detail
And technology, so as not to obscure the understanding of this description.
Similarly, it will be appreciated that in order to simplify the disclosure and help to understand one or more of each inventive aspect, exist
Above in the description of the exemplary embodiment of the present invention, each feature of the invention is grouped together into single implementation sometimes
In example, figure or descriptions thereof.However, the method for the disclosure should be construed to reflect following intention:It is i.e. required to protect
The application claims of shield features more more than the feature being expressly recited in each claim.More precisely, such as right
As claim reflects, inventive aspect is all features less than single embodiment disclosed above.Therefore, it then follows tool
Thus claims of body embodiment are expressly incorporated in the embodiment, wherein the conduct of each claim in itself
The separate embodiments of the present invention.
Those skilled in the art, which are appreciated that, to be carried out adaptively to the module in the equipment in embodiment
Change and they are arranged in one or more equipment different from the embodiment.Can be the module or list in embodiment
Member or component be combined into a module or unit or component, and can be divided into addition multiple submodule or subelement or
Sub-component.In addition at least some in such feature and/or process or unit exclude each other, it can use any
Combination is disclosed to all features disclosed in this specification (including adjoint claim, summary and accompanying drawing) and so to appoint
Where all processes or unit of method or equipment are combined.Unless expressly stated otherwise, this specification (including adjoint power
Profit is required, summary and accompanying drawing) disclosed in each feature can or similar purpose identical, equivalent by offer alternative features come generation
Replace.
Although in addition, it will be appreciated by those of skill in the art that some embodiments described herein include other embodiments
In included some features rather than further feature, but the combination of the feature of be the same as Example does not mean in of the invention
Within the scope of and form different embodiments.For example, in detail in the claims, embodiment claimed it is one of any
Mode it can use in any combination.
The present invention all parts embodiment can be realized with hardware, or with one or more processor run
Software module realize, or realized with combinations thereof.It will be understood by those of skill in the art that can use in practice
Microprocessor or digital signal processor (DSP) come realize in Text region scheme according to embodiments of the present invention some or
The some or all functions of person's whole part.The present invention is also implemented as perform method as described herein one
Divide or whole equipment or program of device (for example, computer program and computer program product).It is such to realize this hair
Bright program can be stored on a computer-readable medium, or can have the form of one or more signal.It is such
Signal can be downloaded from internet website and obtained, and either provided or provided in any other form on carrier signal.
It should be noted that the present invention will be described rather than limits the invention for above-described embodiment, and ability
Field technique personnel can design alternative embodiment without departing from the scope of the appended claims.In the claims,
Any reference symbol between bracket should not be configured to limitations on claims.Word "comprising" is not excluded the presence of not
Element or step listed in the claims.Word "a" or "an" before element does not exclude the presence of multiple such
Element.The present invention can be by means of including the hardware of some different elements and coming real by means of properly programmed computer
It is existing.In if the unit claim of equipment for drying is listed, several in these devices can be by same hardware branch
To embody.The use of word first, second, and third does not indicate that any order.These words can be explained and run after fame
Claim.
Claims (11)
1. a kind of character recognition method, methods described includes:
Using a variety of slit modes, cutting character image obtains a variety of image cutting results respectively, included by each image cutting result
Multiple cutting regions respectively include at least one object to be identified;
Object to be identified in the cutting region counts the synthetic weights weight values of each image cutting result, wherein, it is described comprehensive
Close the system that weighted value includes respectively cutting subregional character features matching degree and shape facility matching degree in described image cutting result
Evaluation;
Optimum image cutting result is screened according to the synthetic weights weight values of each image cutting result;
Recognition result is used as using the corresponding target text in each cutting region in the optimum image cutting result.
2. according to the method described in claim 1, the object to be identified in the cutting region counts each image and cut
The step of synthetic weights weight values of point result, includes:
It is to be identified in the cutting region with recognizing according to subregional object to be identified is respectively cut in described image cutting result
The character features matching degree for the target text that object is obtained, and, respectively cut subregional object to be identified and the target text
The shape facility matching degree of corresponding preset shape, calculates the synthetic weights weight values of described image cutting result.
3. according to the method described in claim 1, described, using a variety of slit modes, cutting character image obtains a variety of figures respectively
Include as the step of cutting result:
Multiple candidate's cut-offs are marked on the character image;
According to different candidate's cut-offs of selection, multiple candidate's cutting point sets are formed respectively;
According to each candidate's cutting point set, character image described in cutting obtains multiple images cutting result respectively.
4. method according to claim 3, the character image includes multiple objects to be slit, described in the word graph
As the upper multiple candidate's cut-offs of mark include following at least one:
Multiple candidate's cut-offs are marked on the impartial multiple positions of character image distance;Or
Search each adjacent but disconnected target object to be slit on the character image, and each target object to be slit it
Between position on mark multiple candidate's cut-offs;Or
Each Object Projection to be slit on the character image is obtained into multiple projection coordinate's points in certain direction reference axis, according to certain
Coordinate corresponding multiple candidate's cuttings of position mark on the character image of projection coordinate's point are not present in the reference axis of direction
Point.
5. according to the method described in claim 1, described, using a variety of slit modes, cutting character image obtains a variety of figures respectively
Include as the step of cutting result:
According to the object to be identified putting in order in the character image, successively to each object to be identified using multiple
Mark window with different label ranges is marked;
Recognize the target text corresponding to the object to be identified that the mark window of different label ranges is marked;
The object to be identified marked according to the mark window and corresponding target text, filter out each object to be identified most
Excellent mark window;
According to the optimal mark window of each object to be identified, character image described in cutting obtains described image cutting result.
6. method according to claim 2, including:
Extract the characteristic vector of the object to be identified in the cutting region;
The word for being matched with the characteristic vector is searched in default characters matching table as the target text;
The cosine value of the object to be identified in the cutting region and the characteristic vector of the target text is calculated, described cut is obtained
Subregional character features matching degree.
7. method according to claim 2, including:
Search preset shape corresponding with the word classification belonging to the target text;
The cosine value of the ratio of width to height of object to be identified and the preset shape in the cutting region is calculated, the cutting is obtained
The shape facility matching degree in region.
8. it is subregional to be identified right respectively to be cut in method according to claim 2, the cutting result according to described image
As the character features matching degree with recognizing the target text that the object to be identified in the cutting region is obtained, and, each cutting
The shape facility matching degree of the object to be identified in region preset shape corresponding with the target text, calculates described image cutting
As a result the step of synthetic weights weight values, includes:
For same image cutting result, multiple average values and shape facility for cutting subregional character features matching degree are calculated
Average value with degree, the character features matching degree average and shape facility matching degree respectively as described image cutting result is equal
Value;
By the character features matching degree average and shape facility matching degree average of described image cutting result and the weight system of distribution
Number is multiplied, and product is summed obtains the synthetic weights weight values of described image cutting result.
9. a kind of character recognition device, described device includes:
Image cutting result acquisition module, for cutting character image to obtain a variety of image cuttings respectively using a variety of slit modes
As a result, multiple cutting regions included by each image cutting result include at least one object to be identified respectively;
Comprehensive weight Data-Statistics module, each image cutting result is counted for the object to be identified in the cutting region
Synthetic weights weight values;The synthetic weights weight values include respectively cutting subregional character features matching degree and shape in described image cutting result
The statistical value of shape characteristic matching degree;
Optimum image cutting result screening module, cuts for screening optimum image according to the synthetic weights weight values of each image cutting result
Divide result;
Recognition result determining module, for being made using the corresponding target text in each cutting region in the optimum image cutting result
For recognition result.
10. a kind of computer equipment, including memory, processor and storage are on a memory and the meter that can run on a processor
Calculation machine program, it is characterised in that any one methods described in claim 1-8 is realized during the computing device described program
The step of.
11. a kind of computer-readable recording medium, is stored thereon with computer program, it is characterised in that the program is by processor
The step of claim 1-8 any one methods describeds are realized during execution.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710257788.8A CN107169496A (en) | 2017-04-19 | 2017-04-19 | A kind of character recognition method and device |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710257788.8A CN107169496A (en) | 2017-04-19 | 2017-04-19 | A kind of character recognition method and device |
Publications (1)
Publication Number | Publication Date |
---|---|
CN107169496A true CN107169496A (en) | 2017-09-15 |
Family
ID=59813317
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710257788.8A Pending CN107169496A (en) | 2017-04-19 | 2017-04-19 | A kind of character recognition method and device |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107169496A (en) |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109145879A (en) * | 2018-09-30 | 2019-01-04 | 金蝶软件(中国)有限公司 | A kind of type fount knows method for distinguishing, equipment and storage medium |
CN109376731A (en) * | 2018-08-24 | 2019-02-22 | 北京三快在线科技有限公司 | A kind of character recognition method and device |
CN109409377A (en) * | 2018-12-03 | 2019-03-01 | 龙马智芯(珠海横琴)科技有限公司 | The detection method and device of text in image |
CN109800697A (en) * | 2019-01-09 | 2019-05-24 | 国网浙江省电力有限公司舟山供电公司 | Transformer target detection and open defect recognition methods based on VGG-net Style Transfer |
CN110222569A (en) * | 2019-05-05 | 2019-09-10 | 北京三快在线科技有限公司 | Method for checking object, device, electronic equipment and readable storage medium storing program for executing |
CN110956058A (en) * | 2018-09-26 | 2020-04-03 | 北京嘀嘀无限科技发展有限公司 | Image recognition method and device and electronic equipment |
CN111340029A (en) * | 2018-12-19 | 2020-06-26 | 富士通株式会社 | Device and method for identifying at least partial address in recipient address |
CN112487848A (en) * | 2019-09-12 | 2021-03-12 | 京东方科技集团股份有限公司 | Character recognition method and terminal equipment |
Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101853126A (en) * | 2010-05-12 | 2010-10-06 | 中国科学院自动化研究所 | Real-time identification method for on-line handwriting sentences |
CN102314616A (en) * | 2010-06-30 | 2012-01-11 | 汉王科技股份有限公司 | Self-adaptation offline handwriting identification method and device |
CN102332096A (en) * | 2011-10-17 | 2012-01-25 | 中国科学院自动化研究所 | Video caption text extraction and identification method |
CN102750534A (en) * | 2012-06-26 | 2012-10-24 | 北京文通科技有限公司 | Method and device for segmenting characters |
CN103984943A (en) * | 2014-05-30 | 2014-08-13 | 厦门大学 | Scene text identification method based on Bayesian probability frame |
-
2017
- 2017-04-19 CN CN201710257788.8A patent/CN107169496A/en active Pending
Patent Citations (5)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN101853126A (en) * | 2010-05-12 | 2010-10-06 | 中国科学院自动化研究所 | Real-time identification method for on-line handwriting sentences |
CN102314616A (en) * | 2010-06-30 | 2012-01-11 | 汉王科技股份有限公司 | Self-adaptation offline handwriting identification method and device |
CN102332096A (en) * | 2011-10-17 | 2012-01-25 | 中国科学院自动化研究所 | Video caption text extraction and identification method |
CN102750534A (en) * | 2012-06-26 | 2012-10-24 | 北京文通科技有限公司 | Method and device for segmenting characters |
CN103984943A (en) * | 2014-05-30 | 2014-08-13 | 厦门大学 | Scene text identification method based on Bayesian probability frame |
Non-Patent Citations (1)
Title |
---|
葛巧瑞: "自然场景下的文字分割及识别研究", 《中国优秀硕士学位论文全文数据库(电子期刊)》 * |
Cited By (13)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN109376731A (en) * | 2018-08-24 | 2019-02-22 | 北京三快在线科技有限公司 | A kind of character recognition method and device |
CN110956058B (en) * | 2018-09-26 | 2023-10-24 | 北京嘀嘀无限科技发展有限公司 | Image recognition method and device and electronic equipment |
CN110956058A (en) * | 2018-09-26 | 2020-04-03 | 北京嘀嘀无限科技发展有限公司 | Image recognition method and device and electronic equipment |
CN109145879B (en) * | 2018-09-30 | 2021-01-12 | 金蝶软件(中国)有限公司 | Method, equipment and storage medium for identifying printing font |
CN109145879A (en) * | 2018-09-30 | 2019-01-04 | 金蝶软件(中国)有限公司 | A kind of type fount knows method for distinguishing, equipment and storage medium |
CN109409377A (en) * | 2018-12-03 | 2019-03-01 | 龙马智芯(珠海横琴)科技有限公司 | The detection method and device of text in image |
CN111340029A (en) * | 2018-12-19 | 2020-06-26 | 富士通株式会社 | Device and method for identifying at least partial address in recipient address |
CN109800697A (en) * | 2019-01-09 | 2019-05-24 | 国网浙江省电力有限公司舟山供电公司 | Transformer target detection and open defect recognition methods based on VGG-net Style Transfer |
CN109800697B (en) * | 2019-01-09 | 2022-10-25 | 国网浙江省电力有限公司舟山供电公司 | Transformer target detection and appearance defect identification method based on VGG-net style migration |
CN110222569A (en) * | 2019-05-05 | 2019-09-10 | 北京三快在线科技有限公司 | Method for checking object, device, electronic equipment and readable storage medium storing program for executing |
CN112487848A (en) * | 2019-09-12 | 2021-03-12 | 京东方科技集团股份有限公司 | Character recognition method and terminal equipment |
WO2021047484A1 (en) * | 2019-09-12 | 2021-03-18 | 京东方科技集团股份有限公司 | Text recognition method and terminal device |
CN112487848B (en) * | 2019-09-12 | 2024-04-26 | 京东方科技集团股份有限公司 | Character recognition method and terminal equipment |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107169496A (en) | A kind of character recognition method and device | |
CN108830209B (en) | Remote sensing image road extraction method based on generation countermeasure network | |
CN107305630B (en) | Text sequence identification method and device | |
CN106529508B (en) | Based on local and non local multiple features semanteme hyperspectral image classification method | |
CN108921166A (en) | Medical bill class text detection recognition method and system based on deep neural network | |
CN103077389B (en) | A kind of combination character level classification and character string level classification text detection and recognition methods | |
US7899249B2 (en) | Media material analysis of continuing article portions | |
Malinverni et al. | Hybrid object-based approach for land use/land cover mapping using high spatial resolution imagery | |
CN110084292A (en) | Object detection method based on DenseNet and multi-scale feature fusion | |
CN106682696A (en) | Multi-example detection network based on refining of online example classifier and training method thereof | |
CN106815604A (en) | Method for viewing points detecting based on fusion of multi-layer information | |
CN107169485A (en) | A kind of method for identifying mathematical formula and device | |
CN111738055B (en) | Multi-category text detection system and bill form detection method based on same | |
JP2013250809A (en) | Multi-class discrimination device, method and program | |
CN103413145A (en) | Articulation point positioning method based on depth image | |
CN103020265A (en) | Image retrieval method and system | |
CN110008899A (en) | A kind of visible remote sensing image candidate target extracts and classification method | |
CN106874901A (en) | A kind of driving license recognition methods and device | |
CN108734159A (en) | The detection method and system of sensitive information in a kind of image | |
Ayhan et al. | Analysis of image classification methods for remote sensing | |
CN106529470A (en) | Gesture recognition method based on multistage depth convolution neural network | |
CN103310205B (en) | A kind of Handwritten Numeral Recognition Method and device | |
CN104732246B (en) | A kind of semi-supervised coorinated training hyperspectral image classification method | |
CN105740901B (en) | Mutative scale object-oriented Classification in Remote Sensing Image antidote based on ontology | |
CN105740877A (en) | Traffic sign recognition method and device, and vehicle |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20170915 |