CN107330429B

CN107330429B - Certificate item positioning method and device

Info

Publication number: CN107330429B
Application number: CN201710349083.9A
Authority: CN
Inventors: 伍更新; 李健; 张连毅; 武卫东
Original assignee: Beijing Sinovoice Technology Co Ltd
Current assignee: Beijing Sinovoice Technology Co Ltd
Priority date: 2017-05-17
Filing date: 2017-05-17
Publication date: 2021-03-09
Anticipated expiration: 2037-05-17
Also published as: CN107330429A

Abstract

The invention provides a method and a device for positioning certificate entries, and relates to the technical field of certificate identification. According to the certificate item positioning method and device, when item positioning is carried out by using the certificate regression model, the actual positions of all items are obtained through regression on the basis of the initial positions of the items in the certificate. Because the initial position of the entry in the certificate is fixed and the standard degree of the certificate image does not influence the initial position of the entry in the certificate, the entry in the certificate is positioned through the certificate regression model, and the accuracy of entry positioning can be improved.

Description

Certificate item positioning method and device

Technical Field

The invention relates to the technical field of certificate identification, in particular to a certificate item positioning method and device.

Background

At present, the application of certificate identification is more and more extensive. The identification of the certificate is to collect various certificate images, such as second-generation identity cards, passports, driving licenses, driving license images and the like, through a mobile terminal, and then scan the certificate images to automatically identify information of each item on the certificate. To enable identification of a certificate, various entries in the certificate need to be located.

In the prior art, when each item in a certificate is located, generally, binarization processing is performed on the certificate to obtain a binary image, and then layout analysis is performed on the binary image to locate the item to be identified.

When the prior art is adopted for item positioning, because the requirement of layout analysis on the standardization of images is higher, the result of the layout analysis can be influenced by the irregular images. The certificate image collected in practical application is usually an irregular image, for example, the collected certificate image often has the problems of uneven illumination, surface reflection, shading on the background and the like. Therefore, in the prior art, when the items to be identified are located through layout analysis, the problem of inaccurate location often occurs.

Disclosure of Invention

In view of the above, the present invention has been made to provide a method and a device for locating a credential item that overcome or at least partially solve the above problems.

According to a first aspect of the invention, there is provided a method of locating a credential entry, the method comprising:

acquiring a certificate area;

determining the position of each item to be identified in the certificate area by using a preset certificate regression model according to the certificate area; the certificate regression model is obtained by training according to the first characteristic points and the second characteristic points; the first feature point is an entry initial position feature point, and the second feature point is an entry actual position feature point.

Optionally, the method further includes:

and training according to the first feature points and the second feature points to obtain the certificate regression model.

Optionally, the step of obtaining the credential regression model by training according to the first feature point and the second feature point includes:

determining a first feature vector according to the first feature point, and determining a second feature vector according to the second feature point;

and training a preset feature algorithm by using the first feature vector and the second feature vector to obtain a certificate regression model.

Optionally, the step of acquiring the certificate region includes:

acquiring a target image;

determining a certificate area by using a preset certificate detection model according to the target image; the certificate detection model is obtained according to sample training; the sample comprises a positive sample and a negative sample, wherein the positive sample is an image containing a certificate area, and the negative sample is an image not containing the certificate area.

Optionally, before the step of determining the certificate region by using a preset certificate detection model according to the target image, the method further includes:

calculating a feature set of each sample based on a feature algorithm; wherein, the feature set comprises at least one feature value corresponding to each sample;

constructing at least one strong classifier by using the feature set of each sample;

and constructing a cascade classifier according to the at least one strong classifier, and determining the cascade classifier as the certificate detection model.

According to a second aspect of the invention there is provided a device for locating a document entry, the device comprising:

the acquisition module is used for acquiring a certificate area;

the determining module is used for determining the position of each item to be identified in the certificate area by using a preset certificate regression model according to the certificate area; the certificate regression model is obtained by training according to the first characteristic points and the second characteristic points; the first feature point is an entry initial position feature point, and the second feature point is an entry actual position feature point.

Optionally, the apparatus further comprises:

and the training module is used for training according to the first characteristic points and the second characteristic points to obtain the certificate regression model.

Optionally, the training module includes:

the first determining submodule is used for determining a first feature vector according to the first feature point and determining a second feature vector according to the second feature point;

and the first training submodule is used for training a preset feature algorithm by using the first feature vector and the second feature vector to obtain a certificate regression model.

Optionally, the obtaining module includes:

the acquisition submodule is used for acquiring a target image;

the second determining submodule is used for determining a certificate area by using a preset certificate detection model according to the target image; the certificate detection model is obtained according to sample training; the sample comprises a positive sample and a negative sample, wherein the positive sample is an image containing a certificate area, and the negative sample is an image not containing the certificate area.

Optionally, the apparatus further comprises:

the calculating submodule is used for calculating a characteristic set of each sample based on a characteristic algorithm; wherein, the feature set comprises at least one feature value corresponding to each sample;

the construction sub-module is used for constructing at least one strong classifier by utilizing the feature set of each sample;

and the third determining submodule is used for constructing a cascade classifier according to the at least one strong classifier and determining the cascade classifier as the certificate detection model.

According to the certificate item positioning method and device provided by the embodiment of the invention, when item positioning is carried out by using a certificate regression model, the actual positions of all items are obtained by regression on the basis of the initial positions of the items in the certificate. Because the initial position of the entry in the certificate is fixed and the standard degree of the certificate image does not influence the initial position of the entry in the certificate, the entry in the certificate is positioned through the certificate regression model, and the accuracy of entry positioning can be improved.

The foregoing description is only an overview of the technical solutions of the present invention, and the embodiments of the present invention are described below in order to make the technical means of the present invention more clearly understood and to make the above and other objects, features, and advantages of the present invention more clearly understandable.

Drawings

Various other advantages and benefits will become apparent to those of ordinary skill in the art upon reading the following detailed description of the preferred embodiments. The drawings are only for purposes of illustrating the preferred embodiments and are not to be construed as limiting the invention. Also, like reference numerals are used to refer to like parts throughout the drawings. In the drawings:

FIG. 1 is a flow chart illustrating steps of a method for locating credential entries according to an embodiment of the present invention;

FIG. 2-1 is a flow chart illustrating steps in another method for locating credential entries provided by embodiments of the present invention;

FIG. 2-2 is a flowchart of certificate region acquisition method steps provided by an embodiment of the invention;

FIGS. 2-3 are flowcharts illustrating steps of a method for obtaining a regression model of a document according to an embodiment of the present invention;

fig. 2-4 are schematic diagrams of a first feature point provided by the embodiment of the invention;

fig. 2-5 are schematic diagrams of a second feature point provided by the embodiment of the invention;

FIG. 3 is a block diagram of a credential entry locator device provided by embodiments of the present invention;

FIG. 4-1 is a block diagram of another credential entry locator device provided by embodiments of the present invention;

fig. 4-2 is a block diagram of another credential entry locator device provided by an embodiment of the invention.

Detailed Description

Exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be embodied in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

Fig. 1 is a flowchart illustrating steps of a method for locating a credential entry according to an embodiment of the present invention, where as shown in fig. 1, the method may include:

step 101, acquiring a certificate area.

The certificate area in the embodiment of the invention is the actual area of the certificate. The certificate can be an identity card, a driving license, a student license and the like, and the embodiment of the invention does not limit the specific types of the certificates. Generally, when a certificate is identified, an image acquisition device acquires an image of the certificate. The captured document image generally includes a document region and a background region. For example, when a document is placed on a table and an image of the document is captured by an image sensor, the captured image of the document includes the document area and a part of the table.

The method and the device have the advantages that the items in the certificate can be conveniently positioned, and meanwhile, the positioning accuracy is improved. When locating an entry in a document, it is necessary to identify the region of the document in the captured image of the document.

And 102, determining the position of each item to be identified in the certificate area by using a preset certificate regression model according to the certificate area.

In the embodiment of the present invention, a certificate regression model can be obtained through training of the first feature point and the second feature point, and then the certificate area obtained in step 101 is input into the certificate regression model to obtain the position of each item to be identified in the certificate area.

The certificate regression model is obtained by training according to a first feature point and a second feature point, wherein the first feature point can be an entry initial position feature point, and the second feature point can be an entry actual position feature point. Because the format of the certificate is fixed, namely the initial position of the item is known, the actual position of the item in the certificate area can be obtained by continuously iterating by using the certificate regression model according to the initial position of the item, and the positioning of the item to be identified in the certificate is completed. And the normalization degree of the certificate image can not influence the initial position of the item in the certificate, so that the positioning accuracy can be improved when the item is positioned through the certificate regression model.

In summary, in the method for locating a certificate entry provided in the embodiments of the present invention, when the certificate regression model is used to locate an entry, the actual position of each entry is obtained by regression based on the initial position of the entry in the certificate. Because the initial position of the entry in the certificate is fixed and the standard degree of the certificate image does not influence the initial position of the entry in the certificate, the entry in the certificate is positioned through the certificate regression model, and the accuracy of entry positioning can be improved.

FIG. 2-1 is a flow chart illustrating steps of another method for locating a credential entry according to embodiments of the present invention, which may include, as shown in FIG. 2-1:

step 201, certificate regions are acquired.

Specifically, fig. 2-2 is a flowchart of steps of a certificate region acquisition method provided in an embodiment of the present invention, and as shown in fig. 2-2, step 201 may include:

and 2011, calculating a feature set of each sample based on a feature algorithm.

The samples in the embodiment of the invention comprise positive samples and negative samples. Wherein the positive swatch can be a document area image and the negative swatch can be an image containing no document area or an image containing a portion of a document area. When sample data is obtained, a certificate image containing a certificate area can be collected from an actual scene, and the certificate area is manually marked to obtain the certificate area as a positive sample. And acquiring a certain number of non-certificate areas as negative samples in a manual screening mode. The certificate image is collected from the actual scene, the obtained sample is more real, and the certificate detection model obtained by training the sample is higher in accuracy rate when the certificate area is obtained. In the embodiment of the invention, the quantity ratio of the positive sample to the negative sample in the sample is not limited, and the quantity ratio of the positive sample to the negative sample can be set according to actual needs. Preferably, the ratio of the two quantities may be 1: 5.

When the negative sample is acquired, the non-certificate area image with the same size as that of the positive sample can be directly acquired as the negative sample, or the size of the negative sample is not considered, and after all the negative samples are acquired, the sizes of all the negative samples and the sizes of the positive samples are set to be the same through normalization processing.

The certificate detection model obtained by sample training can be used for rapidly and accurately determining the certificate area in the certificate image. When the certificate area is determined, an image area with the size consistent with that of the certificate area is randomly selected from the certificate image, then whether the image area is the certificate area or not is judged by using the certificate detection model, and the certificate area is obtained through multiple searches.

Therefore, the certificate area can be abstracted to be the problem of image area classification based on the characteristics, namely, an image area with the size of the certificate area is arbitrarily selected, the image area is classified, and whether the image area belongs to the certificate area or the non-certificate area is judged.

In the embodiment of the present invention, the feature set of each sample may include feature values of multiple features of the sample, that is, one feature set may include multiple feature values. For example, the characteristic of the sample may be brightness, gray scale, or gray scale difference, and the corresponding brightness value, gray scale, or gray scale difference may be used as various characteristic values of the sample.

In an example, based on haar features, in the embodiment of the present invention, a feature set of each sample is calculated by taking gray difference features of the samples as an example, so as to perform schematic explanation. The gray difference features can be divided into three features of upper and lower gray differences, left and right gray differences and 45-degree angle gray differences, and the upper and lower gray difference feature values, the left and right gray difference feature values and the 45-degree angle gray difference feature value of each sample are respectively calculated to obtain a feature set of each sample. For example, an integral graph algorithm may be used to calculate the feature value of each sample.

Thus, a sample has three corresponding feature values, and the three feature values can form a feature set of the sample. In practical applications, the feature set may further include feature values corresponding to other features of the sample, for example, the feature set may further include feature values corresponding to Local Binary Pattern (LBP) features of the sample, which is not limited in this embodiment of the present invention. Meanwhile, the number of eigenvalues included in one feature set can be set according to actual needs, which is not limited in the embodiment of the present invention.

The process of constructing the certificate checking model will be described with 10 samples as an example. Wherein the positive sample is X1, X2, X3, X7, X8 and X9, and the negative sample is X4, X5, X6 and X10.

And respectively calculating the upper and lower gray difference characteristic value, the left and right gray difference characteristic value and the 45-degree angle gray difference characteristic value of each sample in the 10 samples to obtain a characteristic set of each sample. The feature set of each sample comprises an upper gray difference feature value, a lower gray difference feature value, a left gray difference feature value, a right gray difference feature value and a 45-degree angle gray difference feature value. The method comprises the following specific steps: the set of features of X1 may be (X1a, X1b, X1c), the set of features of X2 may be (X2a, X2b, X2c), the set of features of X3 may be (X3a, X3b, X3c), the set of features of X4 may be (X4a, X4b, X4c), the set of features of X5 may be (X5a, X5b, X5c), the set of features of X c may be (X6 c ), the set of X c may be (X7 c ), the set of X c may be (X8 c ), the set of X c may be (X9, X c), the set of features of X c may be (X10, X c, X10).

Step 2012, at least one strong classifier is constructed by using the feature set of each sample.

For example, when constructing a strong classifier, an Adaptive Boosting (Adaptive Boosting) algorithm may be employed. The Adaboost algorithm is an iterative algorithm, and the core idea is to train a plurality of different weak classifiers aiming at the same training set and then assemble the weak classifiers to form a strong classifier. In each iteration training process of the Adaboost algorithm, the sample of the previous weak classifier is reinforced, and the weighted whole sample is used for training the next weak classifier. At the same time, a new weak classifier is added in each round until a certain predetermined sufficiently small error rate is reached or a pre-specified maximum number of iterations is reached.

Since the feature set of a sample includes at least one feature value, at least one strong classifier can be obtained by using the feature set of each sample. For example, the upper and lower gray level difference feature values of each sample are used as a training set, a strong classifier can be constructed, the left and right gray level difference feature values of each sample are used as the training set, a strong classifier can also be constructed, the 45-degree angle gray level difference feature value of each sample is used as the training set, a strong classifier can also be constructed, or a new feature value is added on the basis that the upper and lower gray level difference feature values of each sample are used as the training set, and a strong classifier is reconstructed. The number of the specific strong classifiers can be determined by the number of eigenvalues included in the feature set. Taking three special feature values in the feature set in the above steps as an example, at least three strong classifiers can be constructed by using three feature description values in each sample feature set.

For example, the process of constructing a strong classifier is described by taking the upper and lower gray difference feature values of X1, X2, X3, X4, X5, X6, X7, X8, X9, and X10 as an example of a training set. Assuming that the upper and lower gray level difference characteristic values corresponding to the 10 samples are respectively: x1a ═ 1, X2a ═ 1, X3a ═ 1, X3a ═ 1, X3a ═ 4, X3a ═ 5, X3a ═ 6, X3a ═ 7, X3a ═ 8, and X10a ═ 9, and the 10 feature values can be divided into two categories, one category being "1" indicating a positive sample, i.e., a credential region, and the other category being "-1" indicating a negative sample, i.e., a non-credential region.

Specifically, a strong classifier can be constructed by the following steps:

and step A, initializing weight distribution of each upper and lower gray difference characteristic value.

If there are N samples, the same weight is set for each training sample: 1/N. When weight distribution is initialized for the 10 eigenvalues, the same weight can be set for each eigenvalue: 1/10, let weight W of each sample_mi＝W_1i1/N0.1, where N denotes the number of samples, N10, i 1, 2.

And B, training by utilizing each upper and lower gray difference characteristic value to obtain the weak classifier.

In the embodiment of the invention, a plurality of weak classifiers can be obtained by carrying out iterative training on samples, then a strong classifier is constructed by utilizing the weak classifiers, specifically, 10 samples in a training set are firstly subjected to first iterative training to obtain a first weak classifier, when the iterative training is carried out, a plurality of thresholds can be determined according to characteristic values, the thresholds can be used as classification nodes to classify the characteristic values, and then the weak classifier is determined according to the accuracy of each threshold classification. And finally, calculating the error rate of the weak classifier, and taking the error rate as the coefficient of the weak classifier, wherein the coefficient represents the weight of the weak classifier in the strong classifier. After one round of training is completed, the weight of each sample in the training set can be updated, and when the weight is updated, the weight of the sample which is not accurately classified in the round of training can be improved, and the weight of the sample which is accurately classified can be reduced. And then carrying out a new round of iterative training by using the updated training set to obtain a new weak classifier. After multiple rounds of iterative training are carried out, a plurality of weak classifiers can be obtained, and then a strong classifier can be formed according to the weak classifiers.

Specifically, the training process is as follows:

using a weight distribution D₁The first round of iterative training is performed by the following process:

carrying out a first iteration training on the sample characteristic values, wherein in the iteration, the weight distribution of each sample characteristic value is D₁The weight of each sample feature value is 0.1. In training, first. Determining a threshold value according to the sample characteristic value; then, taking a threshold value as a characteristic value classification node, and classifying the sample characteristic value; and finally, determining the weak classifier according to the classification result.

Specifically, when the threshold v is 2.5, the samples are classified by using v ═ 2.5 as a classification node, samples greater than 2.5 are classified into the "-1" class, and samples less than 2.5 are classified into the "1" class. Since 012 are all less than 2.5, they are classified as "1", while samples "0, 1, 2" themselves correspond to class "1". It can thus be determined that the sample "0, 1, 2" is correctly classified. Since samples "3, 4, 5, 6, 7, 8, 9" are all greater than 2.5, they are classified in the "-1" class, while samples "3, 4, 5, 9" themselves correspond to "-1", and samples "6, 7, 8" themselves correspond to "1". It can thus be determined that sample "3, 4, 5, 9" is correctly classified and sample "6, 7, 8" is misclassified.

Therefore, when the threshold is 2.5, the error rate is 3 × 0.1 to 0.3 (when x <2.5 is 1, and when x >2.5 is-1, the sample "6, 7, 8" is mistaken, and the error rate is 0.3).

When the threshold v is 5.5, classifying the samples by using v ═ 5.5 as a classification node, classifying the samples larger than 5.5 into "-1" class, and classifying the samples smaller than 5.5 into "1" class. Since samples "0, 1,2, 3, 4, 5" are all less than 5.5, they are classified as "1", and samples "0, 1, 2" themselves correspond to class "1", and samples "3, 4, 5" themselves correspond to class "-1". It can thus be determined that the sample "0, 1, 2" is correctly classified and the sample "3, 4, 5" is misclassified. Since samples "6, 7, 8, 9" are all greater than 5.5, they are classified in the "-1" class, whereas samples "6, 7, 8" themselves correspond to a class of "1" and samples "9" themselves correspond to a class of "-1". It can thus be determined that sample "9" is correctly classified and samples "6, 7, 8" are misclassified.

Therefore, when the threshold v is set to 5.5, samples greater than 5.5 are classified as "1", and samples less than 5.5 are classified as "1", the error rate is 6 × 0.1 — 0.6. If the threshold v is taken to be 5.5, samples less than 5.5 are classified as "-1", and samples greater than 5.5 are classified as "1", then samples "0, 1,2, 9" are misclassified, at which point the error rate is 0.4. Therefore, threshold 5.5 corresponds to a minimum error rate of 0.4.

When the threshold v is 8.5, classifying the samples by using v ═ 8.5 as a classification node, classifying the samples larger than 8.5 into "-1" class, and classifying the samples smaller than 8.5 into "1" class. Then the sample "3, 4, 5" is misclassified and the threshold 8.5 corresponds to an error rate of 0.3

Whether the threshold v is 2.5 or 8.5, there are always 3 sample feature values that are mistakenly divided, so that the weak classifier can be determined by selecting one of the threshold 2.5 and the threshold 8.5. In the embodiment of the invention, the threshold value is taken as 2.5 to form a first weak classifier G₁(x)：

Then calculate by G₁(x) The sum of the weights of the misclassified samples is determined as G₁(x) Error rate on training set e₁，e₁＝P(G₁(xi)≠yi)＝3*0.1＝0.3。

Determining calculation G₁(x) E classification error rate of₁Then, based on the classification error rate e₁Calculation of G₁(x) The coefficient represents G₁(x) The weight occupied in the strong classifier. The following is a coefficient calculation formula, a in the coefficient calculation formula_mCoefficient representing weak classifier, e_mThe classification error rate of the weak classifier is represented, m represents the number of current iterations, and log (×) represents a logarithmic function. G can be calculated by the coefficient calculation formula₁(x) The coefficient of (a).

E is to be₁Substituting the above equation for 0.3 yields: a is₁＝0.4236

Then, the weight distribution of the training data is updated for the next iteration. When updating the weights, the weights can be updated by the following formula:

D_m+l＝(W_m+1，1，W_m+l，2，…W_m+l，i…W_m+l，N)

wherein D is_m+1Represents the weight distribution of the m +1 th round, W_m+1，NRepresents the weight, W, of the sample N in the m +1 th round_miRepresents the weight, Z, of the sample i in the mth round_mRepresents a normalization factor, which may be such that D_m+1Becomes a probability distribution, exp (. + -.) represents an exponential function, y_iRepresents the class corresponding to the sample i in the mth round, for example, the class may be 1It may be-1.

Updated weight distribution D₂＝(0.0715,0.0715,0.0715,0.0715,0.0715,0.0715,0.1666,0.1666,0.1666,0.0715)。

Obtaining a classification function f of the weak classifier through a first iteration₁(x)＝a₁*G₁(x)＝0.4236G₁(x)。

The weak classifier is sign (f)₁(x) Sign () represents a classification function, which enables to isolate the sign of the classification function.

Using a weight distribution D₂The second round of iterative training is performed on the training data set, and the process is as follows:

in the iteration of the round, the weight distribution of each sample characteristic value is D₂The weight of each sample feature value is 0.0715,0.0715,0.0715,0.0715,0.0715,0.0715,0.1666,0.1666,0.1666, 0.0715. When the threshold v is 2.5, classifying the samples by using v ═ 2.5 as a classification node, classifying the samples larger than 2.5 into "-1" class, and classifying the samples smaller than 2.5 into "1" class. The sample "6, 7, 8" is misclassified, and the classification error rate at this time is: 3 × 0.1666 ═ 0.4998.

When the threshold v is 5.5, classifying the samples by using v ═ 5.5 as a classification node, classifying the samples larger than 5.5 into the class "1", and classifying the samples smaller than 5.5 into the class "-1". Then the samples "0, 1,2, 9" are misclassified, and the classification error rate at this time is: 4 × 0.0715 ═ 0.286.

When the threshold v is 8.5, classifying the samples by using v ═ 8.5 as a classification node, classifying the samples larger than 8.5 into "-1" class, and classifying the samples smaller than 8.5 into "1" class. The samples "3, 4, 5" are misclassified, with a classification error rate of: 3 × 0.0715 ═ 0.2143.

It can be seen that the classification error rate is the lowest when the threshold v is taken to be 8.5, so the second basic classifier is obtained as:

calculating formula according to classification error rateCan obtain G₂(x) Error rate on training data set e₂＝P(G₂(xi)≠yi)＝0.0715*3＝0.2143。

E is to be₂Substituting the coefficient into the above coefficient calculation formula to obtain a₂＝0.6496

Then, updating the weight distribution of the training data to obtain D₃＝0.0455,0.0455,0.0455,0.1667,0.1667,0.01667,0.1060,0.1060,0.1060,0.0455。

Obtaining the classification function f of the weak classifier through the second iteration₂(x)＝a₁*G₁(x)+a₂*G₂(x)＝0.4236G₁(x)+0.6496G₂(x) Obtaining the weak classifier as sign (f)₂(x))。

Using a weight distribution D₃The training data set is learned, and a third iterative training is carried out, wherein the process is as follows:

in the iteration of the round, the weight distribution of each sample characteristic value is D₃The weight value of each sample feature value is 0.0455,0.0455,0.0455,0.1667,0.1667,0.1667,0.1060,0.1060, 0.0455)

When the threshold v is 2.5, classifying the samples by using v ═ 2.5 as a classification node, classifying the samples larger than 2.5 into "-1" class, and classifying the samples smaller than 2.5 into "1" class. The sample "6, 7, 8" is misclassified, and the classification error rate at this time is: 3 × 0.1060 ═ XXX.

When the threshold v is 5.5, classifying the samples by using v ═ 5.5 as a classification node, classifying the samples larger than 5.5 into the class "1", and classifying the samples smaller than 5.5 into the class "-1". Then the samples "0, 1,2, 9" are misclassified, and the classification error rate at this time is: 3 × 0.0455+0.0715 ═ XXX.

When the threshold v is 8.5, classifying the samples by using v ═ 8.5 as a classification node, classifying the samples larger than 8.5 into "-1" class, and classifying the samples smaller than 8.5 into "1" class. The samples "3, 4, 5" are misclassified, with a classification error rate of: 3 x 0.1667 ═ XXX.

It can be seen that the classification error rate is the lowest when the threshold v is 5.5, so the third basic classifier is obtained as follows:

according to the calculation formula of the classification error rate, G can be obtained₃(x) Error rate on training data set e₃＝P(G₃(xi)≠yi)＝0.0455*4＝0.1820。

E is to be₃Substituting the coefficient into the above coefficient calculation formula to obtain a₃＝0.7514.

Obtaining the classification function f of the weak classifier through the third iteration₃(x)＝a₁*G₁(x)+a₂*G₂(x)+a₃*G₃(x) The weak classifier is sign (f)₃(x) ). Due to sign (f)₃(x) There are 0 misclassification points on the training set and the iterative training process can be ended.

According to the three weak classifiers obtained in the three iterations, a strong classifier g (x) can be constructed.

G(x)＝sign[f₃(x)]

＝sign[a₁*G₁(x)+a₂*G₂(x)+a₃*G₃(x)]

＝sign[0.4236G₁(x)+0.6496G₂(x)+0.7514G₃(x)]

Further, other feature values of each sample can be used as a training set, the steps are repeated, and a strong classifier is constructed. And in the same way, constructing at least one strong classifier by utilizing at least one characteristic value in each sample characteristic set.

Step 2013, constructing a cascade classifier according to the at least one strong classifier, and determining the cascade classifier as the certificate detection model.

For example, the cascaded classifier may be composed of a plurality of strong classifiers. In the embodiment of the invention, a cascade classifier can be constructed according to the strong classifiers constructed in the steps. The cascade classifier is determined as a credential detection model.

In general, the false detection rate of the trained strong classifier is low, but the detection rate is not very high, and if the strong classifier is required to have a high detection rate, the false detection rate thereof is increased accordingly. The detection rate is the ratio of the number of the certificate areas which are accurately detected to the number of the certificate areas contained in the certificate image. The false detection rate is the ratio of the number of non-certificate areas detected as the certificate area to the number of all detected windows in the certificate image. Therefore, if the strong classifier is directly used as the certificate detection model, the certificate area may not be accurately detected, and the certificate detection efficiency is low. Therefore, a cascade classifier can be constructed by utilizing a plurality of strong classifiers to serve as a certificate detection model. By setting the parameters of the cascade classifier, the cascade classifier has higher detection rate and lower false detection rate. Therefore, when the certificate detection model is used for determining the certificate area, the detection efficiency is high.

And step 2014, acquiring a target image.

The target image may be a certificate image acquired when the certificate is detected. For example, if the identification card of "zhang san" of the user is to be identified, at this time, the image acquired by the image acquisition device and containing the "zhang san" identification card is the target image.

Step 2015, determining a certificate area by using a preset certificate detection model according to the target image.

After the target image is acquired, the certificate detection model constructed in step 2013 can be used to determine the certificate area in the target image. Since the image acquired by the image acquisition device includes the certificate region and other regions, for example, the target image including the "zhang san" identification card includes the "zhang san" identification card region and also includes the background region, the identification card region needs to be determined by using the certificate detection model, so that the identification in the subsequent steps is facilitated, and the identification accuracy is improved.

Specifically, the target image can be input into the certificate detection model, when the certificate area is determined, an image area which is the same as the certificate area in size can be selected from the target image through the detection window at will, the certificate detection model judges whether the image area is the certificate area, the whole target image is traversed, and the certificate area is determined. When the target image is traversed through the detection window, random traversal can be performed, that is, an area with the size consistent with that of the certificate area is randomly selected in the target image and judged until the certificate area is determined. And a top-down traversal mode can be adopted, namely, the target image is traversed from the top end of the target image according to the sequence from top to bottom and from left to right, and the traversal is stopped after the certificate area is obtained, so that the subsequent unnecessary traversal operation can be reduced, and the efficiency of determining the certificate area is improved. In the embodiment of the invention, the certificate detection model is constructed, and then the certificate detection model is used for determining the certificate detection model from the target image, so that the certificate area can be rapidly and accurately determined.

It should be noted that, in the step of acquiring the certificate area, the order of each step is not unique, for example, step 2014 may be located before step 2011, which is not limited by the embodiment of the present invention. Meanwhile, the certificate area can be acquired in other manners, which is not limited in the embodiment of the invention.

Step 202, training according to the first feature points and the second feature points to obtain the certificate regression model.

The certificate regression model is obtained by training according to a first feature point and a second feature point, wherein the first feature point can be an item initial position feature point, and the second feature point can be an item actual position feature point. The certificate regression model can locate each item in a certificate region, and the task of item location is to determine the actual position of an item in a given certificate region. In daily life, the distribution of the items of the certificates is almost fixed, that is, the initial position of the item in each certificate is fixed, and due to the deviation existing in the manufacturing process, the actual position of each item in each certificate has a certain deviation from the preset initial position of the item. Thus, item location may be converted into a process of regressing the actual location of the item by the item initial location.

Optionally, fig. 2-3 are flowcharts of steps of a method for obtaining a credential regression model according to an embodiment of the present invention, and as shown in fig. 2-3, step 202 may include:

step 2021, determining a first feature vector according to the first feature point, and determining a second feature vector according to the second feature point.

For example, the first feature point and the second feature point of each sample can be determined by manually labeling the sample of the certificate area. At the time of labeling, the initial position of the entry in each sample can be labeled as a first feature point, and the actual position of the entry can be labeled as a second feature point. For example, the position of the upper left corner of the entry may be labeled, the position of the lower left corner of the entry may be labeled, or both the upper left corner and the lower left corner of the entry may be labeled. Fig. 2 to 4 are schematic diagrams of a first feature point provided by an embodiment of the present invention, as shown in fig. 2 to 4, the initial positions of the items "name", "gender", "nationality", "address", "birth date" and "forensics date" are labeled to show the first feature point of the certificate area, and the black point in fig. 2 to 4 is the first feature point of the "XX certificate". Fig. 2 to 5 are schematic diagrams of a second feature point provided by an embodiment of the present invention, as shown in fig. 2 to 5, the actual positions of the items "name", "gender", "nationality", "address", "birth date" and "field evidence date" are labeled to show the second feature point of the certificate area, and the gray point in fig. 2 to 5 is the second feature point of the "XX certificate".

Then, a first feature vector of the first feature point and a second feature vector of the second feature point of each sample may be determined according to the set feature descriptors. The feature descriptor may be a Scale-invariant feature transform (SIFT) feature or a Histogram of Oriented Gradients (HOG) feature, which is not limited in the embodiment of the present invention.

Specifically, the embodiments of the present invention take SIFT features as examples, which are schematically illustrated. The SIFT feature descriptor can be used for counting the Gaussian image gradient in the neighborhood near the feature point to obtain the feature vector of the feature point. The method is a three-dimensional array, a vector can be obtained by arranging the three-dimensional array according to a certain rule, and the vector is a characteristic vector. Since the feature descriptor is related to the scale of the feature point, the feature vector of the feature point may be determined on the gaussian image corresponding to the feature point. When determining the feature vector according to the SIFT feature descriptor, the neighborhood of the feature point may be first divided into 16 sub-regions.

And rotating the position and the direction of the image gradient in the neighborhood near the feature point by a direction angle theta by taking the feature point as a center. After the positions and the directions of the image gradients in the neighborhood near the feature points are rotated, gradient direction histograms of 8 directions in each sub-region are calculated, and an accumulated value of each gradient direction is drawn to form a seed point. At this time, the gradient direction histogram of each sub-region divides 0 degree to 360 degrees into 8 direction ranges, each of which is 45 degrees. Thus, each sub-region has gradient strength information in 8 directions. Since there are 16 sub-regions, there are 16 × 8 — 128 data in total, and finally a 128-dimensional SIFT feature vector is formed.

After the feature vectors are determined by the SIFT features, the feature vectors may also be subjected to gaussian weighting. By way of example, the weighting process may be performed using a standard gaussian function. The weighting processing is carried out on the feature vector, so that the feature vector can be prevented from being greatly changed due to the small change of the position of the feature point, and a small weight is given to a point far away from the feature point, so that the accuracy of feature vector calculation is ensured.

Step 2022, training a preset feature algorithm by using the first feature vector and the second feature vector to obtain a certificate regression model.

For example, an iterative training may be repeated according to the first feature vector and the second feature vector of each sample by using an SDM (supervisory descending Method) algorithm, and model parameters are calculated to obtain a certificate regression model.

The following description will take 10 certificate region pictures as an example of a sample for constructing a certificate regression model.

With { dⁱAnd f, representing a certificate area picture set, wherein i represents the number of pictures in the certificate area picture set. The picture set may include 10 certificate region pictures, i.e., i ═ 1,2, 3 … 10. With x_*And a second characteristic point representing the certificate area picture i. With x₀The first characteristic point of the certificate area picture i is shown.

To be provided with

P feature point coordinates in the certificate region picture are represented, h is a nonlinear feature extraction function at each feature point, and SIFT features are taken as an example, namely 128-dimensional SIFT features are extracted from each feature point, namely,

the second characteristic point calibrated by manpower is x_*At x_*The extracted SIFT feature can be recorded as phi_*＝h(d(x_*)). Thus, the entry location can be viewed as a feature point from an initial value x₀Convergence to a minimum value x_*The difference Δ x between the feature points is found which minimizes the following function:

f(x₀+△x)＝||h(d(x₀+△x))-Φ_*||²

for example, the feature points may be determined from an initial value x by determining coefficients for iterative training₀Convergence to a minimum value x_*。

By x in certificate area picture₀SIFT feature of (phi)₀As input features, by the objective function:

the coefficient R of the kth iterative training can be determined_kAnd b_kWherein R is_kAnd b_kRepresenting the path of each iteration, and argmin (x) represents the independent variable value when the function value is minimum, and the parameters of each iteration training are usedAnd determining the certificate regression model.

And 203, determining the position of each item to be identified in the certificate area by using a preset certificate regression model according to the certificate area.

At the time of actual recognition, the credential regions can be entered into the credential regression model, since the format of the credential is fixed, i.e., the entry initial position is known. Therefore, the certificate regression model can be utilized to continuously iterate according to the initial position of the item, so that the actual position of the item in the certificate area can be obtained, and the positioning of the item to be identified in the certificate is completed. Because the normalization degree of the certificate image can not influence the initial position of the item in the certificate, the positioning accuracy can be improved when the item is positioned through the certificate regression model.

Meanwhile, the certificate item positioning method provided by the embodiment of the invention can be completed through the certificate regression model when the item to be identified in the certificate area is positioned, and compared with the positioning method which can be completed through a plurality of steps in the prior art, the method simplifies the item positioning operation and improves the item positioning efficiency.

In summary, the method for locating a certificate entry provided in the embodiments of the present invention determines a certificate detection model from a target image by constructing the certificate detection model, quickly and accurately determines a certificate area, and then locates an entry by using a certificate regression model, to determine a position of each entry to be identified in the certificate area. When the certificate regression model is used for item positioning, the actual positions of all items are obtained through regression on the basis of the initial positions of the items in the certificate. Because the initial position of the entry in the certificate is fixed and the standard degree of the certificate image does not influence the initial position of the entry in the certificate, the entry in the certificate is positioned through the certificate regression model, and the accuracy of entry positioning can be improved.

FIG. 3 is a block diagram of a credential entry locator device according to an embodiment of the present invention, and as shown in FIG. 3, the device 30 may include:

an acquisition module 301 for acquiring a credential area.

A determining module 302, configured to determine, according to the certificate region, a position of each entry to be recognized in the certificate region by using a preset certificate regression model; the certificate regression model is obtained by training according to the first characteristic points and the second characteristic points; the first feature point is an entry initial position feature point, and the second feature point is an entry actual position feature point.

In summary, in the credential item positioning device provided in the embodiment of the present invention, when the credential regression model is used to perform item positioning, the actual position of each item is obtained by regression based on the initial position of the item in the credential. Because the initial position of the entry in the certificate is fixed and the standard degree of the certificate image does not influence the initial position of the entry in the certificate, the entry in the certificate is positioned through the certificate regression model, and the accuracy of entry positioning can be improved.

FIG. 4-1 is a block diagram of another credential entry locator device provided by embodiments of the invention, and as shown in FIG. 4, the device 40 can include:

an acquisition module 401 for acquiring a credential area.

A determining module 402, configured to determine, according to the certificate region, a position of each entry to be identified in the certificate region by using a preset certificate regression model; the certificate regression model is obtained by training according to the first characteristic points and the second characteristic points; the first feature point is an entry initial position feature point, and the second feature point is an entry actual position feature point.

Fig. 4-2 is a block diagram of another credential entry locator device provided by an embodiment of the invention, and as shown in fig. 4-2, the device 40 can include: an acquisition module 401, a determination module 402, a training module 403,

an acquisition module 401 for acquiring a credential area.

And a training module 403, configured to train to obtain the credential regression model according to the first feature point and the second feature point.

Optionally, the training module 403 may include:

Optionally, the obtaining module 401 may include:

the acquisition submodule is used for acquiring a target image;

Optionally, the obtaining module 401 may further include:

The embodiments in the present specification are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.

As is readily imaginable to the person skilled in the art: any combination of the above embodiments is possible, and thus any combination between the above embodiments is an embodiment of the present invention, but the present disclosure is not necessarily detailed herein for reasons of space.

In the description provided herein, numerous specific details are set forth. It is understood, however, that embodiments of the invention may be practiced without these specific details. In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Similarly, it should be appreciated that in the foregoing description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof for the purpose of streamlining the invention and aiding in the understanding of one or more of the various inventive aspects. However, the disclosed method should not be interpreted as reflecting an intention that: that the invention as claimed requires more features than are expressly recited in each claim. Rather, as the following claims reflect, inventive aspects lie in less than all features of a single foregoing disclosed embodiment. Thus, the claims following the detailed description are hereby expressly incorporated into this detailed description, with each claim standing on its own as a separate embodiment of this invention.

Those skilled in the art will appreciate that the modules in the device in an embodiment may be adaptively changed and disposed in one or more devices different from the embodiment. The modules or units or components of the embodiments may be combined into one module or unit or component, and furthermore they may be divided into a plurality of sub-modules or sub-units or sub-components. All of the features disclosed in this specification (including any accompanying claims, abstract and drawings), and all of the processes or elements of any method or apparatus so disclosed, may be combined in any combination, except combinations where at least some of such features and/or processes or elements are mutually exclusive. Each feature disclosed in this specification (including any accompanying claims, abstract and drawings) may be replaced by alternative features serving the same, equivalent or similar purpose, unless expressly stated otherwise.

Furthermore, those skilled in the art will appreciate that while some embodiments described herein include some features included in other embodiments, rather than other features, combinations of features of different embodiments are meant to be within the scope of the invention and form different embodiments. For example, in the claims, any of the claimed embodiments may be used in any combination.

It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The usage of the words first, second and third, etcetera do not indicate any ordering. These words may be interpreted as names.

Claims

1. A method of locating a credential entry, the method comprising:

acquiring a document area, comprising: acquiring a target image, selecting an image area with the size consistent with that of a certificate area in the target image through a detection window, judging whether the image area is the certificate area or not by using a preset certificate detection model, and traversing the whole target image until the certificate area is determined; wherein the traversal is a random traversal or a top-down traversal;

2. The method of claim 1, wherein the step of obtaining the credential regression model by training based on the first feature points and the second feature points comprises:

3. The method of claim 1, wherein the document detection model is obtained from sample training; the sample comprises a positive sample and a negative sample, wherein the positive sample is an image containing a certificate area, and the negative sample is an image not containing the certificate area.

4. The method of claim 3, wherein prior to the step of determining a document region from the target image using a preset document detection model, the method further comprises:

5. A device for locating a credential item, the device comprising:

the acquisition module is used for acquiring a certificate area;

the determining module is used for determining the position of each item to be identified in the certificate area by using a preset certificate regression model according to the certificate area; the certificate regression model is obtained by training according to the first characteristic points and the second characteristic points; the first feature point is an entry initial position feature point, and the second feature point is an entry actual position feature point;

wherein, the obtaining module includes:

the acquisition submodule is used for acquiring a target image;

and the second determining submodule is used for selecting an image area with the size consistent with that of the certificate area in the target image through the detection window, judging whether the image area is the certificate area or not by using a preset certificate detection model, and traversing the whole target image until the certificate area is determined.

6. The apparatus of claim 5, wherein the training module comprises:

7. The apparatus of claim 5,

wherein the certificate detection model is obtained according to sample training; the sample comprises a positive sample and a negative sample, wherein the positive sample is an image containing a certificate area, and the negative sample is an image not containing the certificate area.

8. The apparatus of claim 7, wherein the obtaining module further comprises: