KR20170089745A - Method and apparatus for positioning key points - Google Patents

Abstract

Provided is a feature point positioning apparatus which comprises: a feature extracting unit for extracting a nonlinear feature of an image; an updating unit for repeatedly updating a shape coefficient based on the nonlinear feature; and a position detecting unit for detecting a position of a feature point of the image based on a statistical shape model obtained by training with the shape coefficient after updating.

Description

METHOD AND APPARATUS FOR POSITIONING KEY POINTS [0002]

It is related to computer vision technology. And more particularly to a method and apparatus for minutia positioning.

Positioning of facial features is a typical problem in the visual field of computers, and has already had a research history of more than 20 years. Positioning of facial feature points is a difficult problem due to various indeterminate factors such as face posture, facial expression, and illumination. There are several algorithms in a relatively typical human facial feature point positioning algorithm, and recent feature point positioning techniques are making great strides.

However, such an existing algorithm may not be suitable to be implemented on a mobile platform. In order to be freely used on the mobile platform, it is necessary to study algorithms that can reduce the size of the model file and increase the precision and speed at the same time.

The feature point positioning apparatus according to one embodiment includes a feature extraction unit for extracting a nonlinear feature of a detected image, an update unit for performing a repetitive update on the feature coefficient based on the training-acquired regression factor matrix and the nonlinear feature, And a position detector for detecting a position of a feature point of the detected image based on a statistical shape model obtained by tracing a shape coefficient.

According to an embodiment, the update unit may obtain the similarity transformation matrix by aligning the intermediate shape and the average shape acquired through the update, and proceed the similarity transformation on the intermediate shape and the detected image based on the similarity transformation matrix have.

According to an exemplary embodiment of the present invention, the update unit may perform a process of aligning an average shape and an intermediate shape obtained through an update using a multiresolution pyramid frame, performing a minutia positioning from a first resolution image to a threshold, A similarity transformation matrix can be obtained by mapping feature points to an image having a second resolution larger than the first resolution.

The update unit may invert the feature points acquired based on the similarity transformation matrix obtained through the iterative update to the coordinate system of the detected image.

는 타겟 함수

를 통하여 획득하고,

는 k번째 반복 후의 제i번째 샘플 형태일 수 있다.The update unit may update the similarity transformation matrix < RTI ID = 0.0 >

Target function

Lt; / RTI >

May be in the form of the ith sample after the kth iteration.

The updating unit according to an embodiment may determine a regression factor matrix based on the optimal shape coefficient update amount of the training sample and the nonlinear characteristic.

The updating unit may learn an average texture feature in the training sample and determine a difference value between the nonlinear texture feature extracted from the training sample and the average texture feature as nonlinear characteristics of the training sample.

The nonlinear texture feature according to one embodiment comprises a combination of one kind of nonlinear texture feature or at least two types of nonlinear texture features and is capable of generating a similar or different nonlinear texture feature or at least two non- Combinations can be used.

The updating unit according to an embodiment may determine a regression factor matrix by minimizing an error between an optimal shape factor update amount and a shape factor update amount calculated based on the nonlinear feature.

는 타겟 함수

를 통하여 얻고, 여기서, N은 트레이닝 샘플의 총수를 의미하고,

는 k번째 반복 때, i번째 샘플의 최적 형태 계수 업데이트 량이고,

는 k번째 반복 때, i번째 샘플의 비선형 특징이며,

는 k번째 반복 때, i번째 샘플의 비선형 특징에 기초하여 계산하여 획득한 형태 계수 업데이트 량일 수 있다.In a feature point positioning apparatus according to an embodiment, a regression factor matrix

Target function

, Where N is the total number of training samples,

Is the optimal shape coefficient update amount of the i < th > sample at the k <

Is a non-linear characteristic of the i-th sample at the k-th iteration,

May be a form factor update amount calculated and obtained based on the nonlinear characteristic of the i-th sample at the k-th repetition.

A feature point positioning method according to an embodiment includes extracting a nonlinear feature of a detected image, repeatedly updating a feature coefficient based on a regression factor matrix obtained through training and the nonlinear feature, And detecting the feature point of the detected image based on the statistical shape model obtained by training.

In a minutia positioning method according to an embodiment, the step of repetitively updating the shape coefficient comprises: obtaining a similarity transformation matrix by aligning an average shape and an intermediate shape obtained through an update; and calculating a similarity transformation matrix based on the similarity transformation matrix And performing a similarity transformation on the intermediate image and the detected image.

The method of claim 1, wherein obtaining the similarity transformation matrix comprises: aligning an average shape and an intermediate shape acquired through an update using a multiresolution pyramid frame; And positioning the feature point by mapping the result of the positioning of the feature point to the image of the second resolution larger than the first resolution.

The method may further include the step of inversely transforming the feature points acquired on the basis of the similarity transformation matrix obtained through the iterative update to the coordinate system of the detected image, in the feature point positioning method according to an exemplary embodiment.

는 타겟 함수

를 통하여 획득되고, 는 k번째 반복 후의 제i번째 샘플 형태일 수 있다.In a minutiae positioning method according to an embodiment, the similarity transformation matrix used in the iterative updating of the k-

Target function

And may be in the form of the ith sample after the kth iteration.

In a feature point positioning method according to an embodiment, determining a nonlinear characteristic of a training sample and obtaining a regression factor matrix based on the optimal shape factor update amount of the training sample and the nonlinear characteristic may be included.

The feature point positioning method according to one embodiment, wherein the nonlinear feature comprises a nonlinear texture feature, and wherein determining a nonlinear feature of the training sample comprises learning an average texture feature in the training sample, And determining the difference between the extracted nonlinear texture feature and the average nonlinear texture feature as nonlinear characteristics of the training sample.

In a feature point positioning method according to an embodiment, the nonlinear texture feature may comprise a combination of one or more types of nonlinear texture features, and may use a combination of the one or more nonlinear texture features when repeating the update.

Determining a regression factor matrix based on an optimal shape factor update amount of the training sample and the nonlinear feature comprises calculating an optimal shape factor update amount and a shape calculated based on the nonlinear feature And determining a regression factor matrix by minimizing the error between the coefficient update amount.

는 타겟 함수

를 통하여 획득되고, N은 트레이닝 샘플의 총수를 의미하고,

는 k번째 반복 때, i번째 샘플의 최적 형태 계수 업데이트 량이고,

는 k번째 반복 때, i번째 샘플의 비선형 특징이며,

는 k번째 반복 때, i번째 샘플의 비선형 특징에 기초하여 계산하여 획득한 형태 계수 업데이트 량일 수 있다.In a minutiae positioning method according to an embodiment, the regression factor matrix used for the k < th >

Target function

N is the total number of training samples,

Is the optimal shape coefficient update amount of the i < th > sample at the k <

Is a non-linear characteristic of the i-th sample at the k-th iteration,

May be a form factor update amount calculated and obtained based on the nonlinear characteristic of the i-th sample at the k-th repetition.

1 is a schematic diagram of a key point positioning device according to one embodiment.
2A is an overall flow diagram of a feature point positioning method according to one embodiment.
FIG. 2B is a view for explaining a minutia point positioning method according to an embodiment.
3 is an illustration of a statistical model in accordance with one embodiment.
Figures 4A and 4B are an illustration of an initial shape of a feature point at a single resolution and a shape after a first iteration according to an embodiment.
5 is a diagram for explaining a repetitive update according to an embodiment.
FIG. 6 is a statistical model coefficient regression (SMCR) algorithm according to an embodiment.
FIG. 7 is a flowchart of a multi-resolution statistical model coefficient regression algorithm incorporating the algorithms of FIGS. 4-6.

Specific structural or functional descriptions of embodiments are set forth for illustration purposes only and may be embodied with various changes and modifications. Accordingly, the embodiments are not intended to be limited to the particular forms disclosed, and the scope of the disclosure includes changes, equivalents, or alternatives included in the technical idea.

The terms first or second, etc. may be used to describe various elements, but such terms should be interpreted solely for the purpose of distinguishing one element from another. For example, the first component may be referred to as a second component, and similarly, the second component may also be referred to as a first component.

It is to be understood that when an element is referred to as being "connected" to another element, it may be directly connected or connected to the other element, although other elements may be present in between.

The singular expressions include plural expressions unless the context clearly dictates otherwise. In this specification, the terms "comprises ", or" having ", and the like, are used to specify one or more of the described features, numbers, steps, operations, elements, But do not preclude the presence or addition of steps, operations, elements, parts, or combinations thereof.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by one of ordinary skill in the art. Terms such as those defined in commonly used dictionaries are to be interpreted as having a meaning consistent with the meaning of the context in the relevant art and, unless explicitly defined herein, are to be interpreted as ideal or overly formal Do not. Hereinafter, embodiments will be described in detail with reference to the accompanying drawings. Like reference symbols in the drawings denote like elements.

As technology advances, existing mobile devices, especially mobile phones and tablet PCs, are equipped with front-facing cameras. In hardware, it is possible to conveniently position the feature points of the human face through the front camera. At the same time, existing mobile phones and tablet PCs have high computational power, which can make complex image processing algorithms hardware-enabled.

This paper proposes a feature point positioning algorithm with small size, high accuracy, and high speed for the mobile platform. The basic idea is to first train one statistical form model offline, extract the nonlinear characteristics of the image to be detected, and perform a repetitive update on the form factor based on the regression factor matrix obtained by training and the nonlinear characteristic, And the feature point position of the image to be detected is detected based on the statistical shape model obtained by training and the shape coefficient after. The present invention can train the statistical morphological model in different types of models in advance and detect the minutiae of the corresponding morphology so that the technical solution provided by the present invention is not limited to the detection of the minutiae of the human face, Site, ultrasound image, and the like. In the present application, only the minutiae positioning of a human face will be described as an example.

For example, the statistical form model trained can represent all normal human face shapes with only ten left and right shape coefficients. Compared to a complex model, it allows a simple depiction of the shape of a human face by reducing dimensions through relatively small coefficients. As a result, human facial feature point positioning results in finding the optimal shape factor. It is possible to acquire the optimal shape factor through repetitive update of the shape factor using the method of extracting the nonlinear feature by linear regression, and obtain the position of each feature point based on the statistical shape model obtained by training offline. In this process, the number of rows in the regression factor matrix is equal to the number of coefficients in the statistical model and is not related to the number of minutiae. The number of rows in the regression factor matrix is less than one-ninth of the traditional SDM algorithm. As a result, the size of the model file is greatly reduced, the speed is improved, and the positioning accuracy is not reduced.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a diagram of a feature point positioning apparatus according to one embodiment. FIG.

1, the system includes a feature extraction unit 120, an update unit 130, and a position detection unit 140, and further includes a training unit 150, a storage unit 160, and an image collection unit 110 Which is indicated by a dotted line in the drawing. The system includes a feature extraction unit 120 for extracting a nonlinear feature of an image to be detected, an update unit 130 for performing a repetitive update on a feature coefficient based on the trained and obtained nonlinear feature, And a position detector 140 for detecting the position of the feature point of the image to be detected based on the updated shape factor and the statistical model obtained by training. The training unit 150 learns the statistical form module by using the image of the marked minutia position as a training sample, and the determination of the non-linear characteristic of the training sample is performed based on the update amount of the optimization form coefficient of the training sample and the non- The storage unit 160 may store the statistical form module and other necessary information obtained by training, and the image collecting unit 110 may collect the video and / or image inclusions.

Corresponding to the apparatus of FIG. 1, the present invention provides a method of minutia positioning, comprising: extracting non-linear features of an image to be detected; Repeatedly updating the form factor based on the regression factor matrix obtained by training and the nonlinear characteristic; And detecting the feature point position of the image to be detected based on the updated shape factor and the statistical shape model obtained by training.

2A is an overall flow diagram of a method for positioning a human face feature point of the present invention and includes an off-line training portion and an on-line alignment portion. Among them, the offline training part includes the following steps.

In step 211, a human face image in which the marked minutiae are located is input. In step 212, the statistical model is trained using the human face image in which the marked minutiae are located as training samples. In step 213, the statistical texture model is trained using the face image in which the marked minutiae are located as training samples, and steps 211 to 213 may be selectively performed. In step 214, at least two initial values are generated for each sample to increase the training samples. In step 215, it is determined whether the update is first-time iterated. If the update is the first iteration, The nonlinear feature is extracted and passed to step 216. Otherwise, the previously updated intermediate form is aligned with the average shape and an affine transformation is performed, and a nonlinear feature is performed around the feature point on which the affine transformation is performed (217). In step 218, the statistical form module coefficient update amount for the current form and the actual form is calculated. In step 219, the linear shape coefficient regression factor matrix is calculated based on the update amount of the shape model coefficient and the nonlinear characteristic, the shape factor update amount is calculated based on the linear shape coefficient regression factor matrix and the nonlinear characteristic acquired in the previous step , The current form is updated using the update amount to update the current form. Steps 215 to 219 are repeated until the designated order is converged or circulated, and then the step 220 is performed to output a regression factor matrix, a morphological model, an average texture, and the like.

The on-line positioning portion in Fig. 2a performs the following steps.

In step 231, a video including a human face of a width, a human facial envelope frame, and an offline model are input. In step 232, a human face area is cut out from the video, the size is reduced, Initialize. In step 233, it is determined whether or not the update is to be performed for the first time. If the first iteration is performed, the nonlinear feature is extracted around the current feature point and transmitted to the step 235. Otherwise, Align 234 to affine transform the image and extract nonlinear features around the next current feature point. In step 235, a form factor update quantity is calculated (236) based on the linear shape coefficient regression factor matrix and the nonlinear feature obtained with offline training, and the form factor is updated with this form factor update quantity to update the current form 237). Steps 233 to 237 are repeated to obtain the final human facial feature point position when the current form is converted into the original input image coordinate system.

FIG. 2B is a key step explanatory view of the human face feature point positioning method of the present invention, and includes the offline training part and the on-line positioning part in the same manner. Among them, the offline training part includes the following steps.

Step 1: Learn the statistical model by the following formula.

은 평균형태,

는 국부형태 계수,

는 기초벡터를 의미한다.here,

The average shape,

Is the local form factor,

Means a base vector.

Step 2: Learn through minimizing the equation below.

는 최적형태 모델 계수의 업데이트 량이고;

는 표기한 진실형태가 대응하는 형태 계수며;

는 k차 반복 후의 형태 계수고,

는 영상I의 현재 형태

에서 추출한 비선형 특징이다.here,

Is the update amount of the optimal shape model coefficient;

Is the corresponding form factor for the truth form represented;

Is the shape coefficient after k-th iteration,

The current form of the video I

Is a nonlinear feature extracted from

Step 3: Update the form factor according to the formula below.

Step 4: Repeat steps 2 and 3 until the convergence or iteration order reaches the specified order and then terminate.

The online alignment portion shown in FIG. 2B includes the following steps.

Step 1: Initialize the shape of the face area by extracting the human face area from the original image. For example, in step 1, the drawing 201 is a raw image, and the drawing 202 located on the right is a human face after initialization.

Step 2: The form coefficients are repeatedly updated until convergence, as shown in the diagrams (203) to (206), according to the following equation.

Step 3: The human face form is synthesized according to the following equation and inverse transformed to the input image coordinate system to finally obtain the figure 207 of FIG. 2B.

The entire flow of the above algorithm is introduced, and details of some of them are described below.

1. Training of statistical model

After removing the global similarity transformation elements such as rotation, contraction, and translation for normal human face forms, the conversion of human face forms is finite in practice and there are not many major change patterns. The present invention can display human face shapes using a more precise method, that is, display the shapes of various human faces with the use of fewer coefficients. The present invention is the reason for requiring the training of the statistical model.

In order to train a precise statistical model, it is first necessary to standardize the input face type. Procrustes Analysis, Rotation, Reduction, etc., and the specific implementation steps are described in the related article: T.F. Cootes and C. J. Taylor, Statistical Models of Appearance for Computer Vision, 2004.

과 한 개 조의 기본 벡터

를 획득하고, 동시에 전체적 유사성 변환을 실현하기 위하여 추가적인 4개의 기본 벡터

의 부가가 필요할 수 있다. 평균 형태에서 서로 다른 가중치의 겹침을 통하여 하나의 특정된 사람얼굴 형태를 생성할 수 있고, 전체적 유사성 변환을 추가할 수 있으며, 이는 아래의 공식으로 표시할 수 있다.After the standardization process, the average shape (PCA)

And a set of basic vectors

And at the same time an additional four basic vectors < RTI ID = 0.0 >

May be required. Through the overlap of different weights in the mean form, one specific human face form can be created, and a global similarity transformation can be added, which can be expressed by the following formula.

는 국부형태 계수고,

는 전체적 유사성 변환 계수며, N(,)은 한 개 점(x, y)에 대한 전체적 유사성 변환을 나타낸다. here,

Is the local form factor,

Is the global similarity transform coefficient, and N (,) represents the global similarity transform for a single point (x, y).

In particular, the overall similarity transformation for the mean form can be expressed as:

는 각각

,

,

,

이다.Here, the four basic vectors

Respectively

,

,

,

to be.

에 대응되는 도면(302)은 평균적인 형태이고 도면(303)은 앞 세 개의 기본 벡터를 각각 평균 형태에 겹쳐 표시한 영상(도에서 화살표를 가진 작은 선단은 기본 벡터를 표시함)이다. 첫 번째 기본 벡터(303)는 사람 얼굴이 좌우로 회전 할 때의 형태 변화를 묘사하였고, 두 번째 기본 벡터(304)는 사람 얼굴이 상하로 운동할 때의 형태 변화이며, 세 번째 기본 벡터(305)는 살찐 얼굴과 마른 얼굴 사이의 형태 차이를 묘사하였다. 도 3이 도시한 예시는 단지 기본 벡터가 묘사할 수 있는 정보를 예를 들어 설명하였고, 실제응용에서, 기본 벡터가 묘사할 수 있는 정보는 상기 예제에 한정되지 않는다.3 is an illustration of one statistical model in accordance with one embodiment. The input 301 is an image of several human face forms superimposed together after normalization, and as can be seen, the distribution of points is close to the Gaussian distribution. From the left to the second drawing,

(302) is an average shape, and the figure (303) is an image in which the three preceding basic vectors are superimposed on the average shape, respectively. The first basic vector 303 describes a morphological change when the human face rotates from side to side. The second basic vector 304 is a morphological change when the human face moves up and down. The third basic vector 305 ) Described the difference in shape between a fat face and a dry face. The example shown in FIG. 3 illustrates only information that the basic vector can describe, and in practical applications, the information that the basic vector can describe is not limited to the above example.

When specifying a human face shape, the local form factor and the global similarity transform coefficient can be obtained through the following formula.

은 고정형태에 대하여 전체적 유사성 변환을 하는 역변환을 나타내고, 이는 형태 표준화의 조작과 유사하다. here,

Represents an inverse transformation that performs a global similarity transformation on a fixed form, which is similar to the manipulation of shape normalization.

2. Statistical form model. Regression-based human face sorting method

Statistical morphology model The basic idea of the human face alignment method based on the regression is to extract nonlinear features around each current feature point, combine the nonlinear features into one vector, and use this nonlinear feature to update the statistical model coefficient .

는 현재 형태(k번째 반복 후의 형태)가 대응하는 형태 계수고,

는 현재 형태에서 추출한 비선형 특징이며,

는 k번째 반복 업데이트 후에 사용한 회귀인자이다. 설명의 편의를 위하여 국부 형태 계수와 전체적 유사성 변환계수를 합친 것을

로 표기한다.here,

Is a form coefficient corresponding to the current form (the form after the k-th iteration)

Is a nonlinear feature extracted from the current form,

Is the regression factor used after the kth iteration update. For convenience of explanation, it is assumed that the sum of the local form factor and the global similarity transform coefficient

.

이 획득될 것이 요구된다. 본 발명에서,

는 아래 타겟 함수의 최소화를 통하여 획득할 수 있다.For accurate work, we use regression coefficients

Is required to be obtained. In the present invention,

Can be obtained by minimizing the target function below.

는 k번째 반복 때, i번째 샘플의 최적 형태 계수 업데이트 량(즉 실제의 형태 계수와 현재 형태 계수의 차)이고,

는 k번째 반복 때, i번째 샘플의 비선형 특징이다. 상기 식은 아래의 최소 제곱법이 있다.In the equation, N means the total number of training samples,

(I.e., the difference between the actual shape coefficient and the current shape coefficient) of the i-th sample at the k-th iteration,

Is a non-linear feature of the ith sample at the kth iteration. This equation has the following least squares method.

In addition to the target function and solution formulas presented here, other target functions and solution formulas can be used. Illustratively, ridge regression is employed to replace linear regression, and nonlinear features are projected into the space of the texture model . See Table 1 for specific methods.

Target function and solution formula of SMCR algorithm

Based on the same code frame, we can also compare it with the supervisory descent algorithm and the results are shown in Table 2. In the comparison results, the accuracy of the statistical model coefficient regression algorithm on both LFPW and iBUG databases is higher than that of the supervisory descent algorithm.

Comparison of statistical model coefficient regression algorithm and supervisory descent algorithm

Statistical model coefficient regression algorithm

Supervision descent algorithm: Statistical model It is very similar to the regression algorithm, but it replaces p only x.

3. Rearrangement of intermediate results

What we found in the actual training and testing is that the form obtained through the first or second iteration is already close to the actual form. Referring to FIGS. 4A and 4B, FIG. 4A is an initial form and FIG. 4B is a form after a first repetition. If there is a constant rotation on the human face of the image, even if the current shape already detects the rotation, it may not be able to correct the face image when extracting nonlinear features in subsequent iterations.

A flow chart of the person sorting algorithm that rearranges the intermediate results is shown in Fig.

를 찾아서, 하기 타겟 함수를 최소화하도록 한다.On the basis of the above observation, the present invention is characterized in that, when the iterative update is performed each time, the current face shape and the average face shape are aligned, and one global similarity transformation matrix

To minimize the following target function.

는 제k차 반복 후 i번째 샘플의 형태이다. 전체적 유사성 변환행렬을 획득한 후, 현재 형태와 영상에 대하여 전체적 유사성 변환을 하고, 이에 기초하여 다음 번 반복을 수행한다(511~515). 이로써 회전과 축소로 생기는 불리한 영향을 제거할 수 있어, 추출된 비선형 특징이 특징점 위치 편이를 더 잘 반영할 수 있다.here,

Is the form of the i-th sample after the k-th iteration. After obtaining the global similarity transformation matrix, the global similarity transformation is performed on the current form and the image, and the next iteration is performed based on the similarity transformation (511-515). This can eliminate the adverse effects of rotation and contraction, and the extracted nonlinear features can better reflect feature location shifts.

4. Multi-resolution extension

6 is an exemplary diagram of an extension to multiple resolutions according to one embodiment.

Multi-resolution frames are again used as a common means in each algorithm for equilibrium of precision and speed. In the human face algorithm to which the present invention is associated, the same multi-resolution frame can be employed. First, the human facial feature point is positioned (601) on a relatively low resolution image, and the result is mapped to a high resolution image to continuously perform repetitive precision positioning (602).

Finally, FIG. 7 is a flow diagram of a multi-resolution SMCR algorithm including reordering intermediate results. The diagrams 711 to 715 in FIG. 7 show repetitive update performances in the short resolution described above in detail, and 721 to 722 denote performing repetitive positioning in multiple resolutions.

The embodiments described above may be implemented in hardware components, software components, and / or a combination of hardware components and software components. For example, the devices, methods, and components described in the embodiments may be implemented within a computer system, such as, for example, a processor, a controller, an arithmetic logic unit (ALU), a digital signal processor, such as an array, a programmable logic unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions. The processing device may execute an operating system (OS) and one or more software applications running on the operating system. The processing device may also access, store, manipulate, process, and generate data in response to execution of the software. For ease of understanding, the processing apparatus may be described as being used singly, but those skilled in the art will recognize that the processing apparatus may have a plurality of processing elements and / As shown in FIG. For example, the processing unit may comprise a plurality of processors or one processor and one controller. Other processing configurations are also possible, such as a parallel processor.

The software may include a computer program, code, instructions, or a combination of one or more of the foregoing, and may be configured to configure the processing device to operate as desired or to process it collectively or collectively Device can be commanded. The software and / or data may be in the form of any type of machine, component, physical device, virtual equipment, computer storage media, or device , Or may be permanently or temporarily embodied in a transmitted signal wave. The software may be distributed over a networked computer system and stored or executed in a distributed manner. The software and data may be stored on one or more computer readable recording media.

The method according to an embodiment may be implemented in the form of a program command that can be executed through various computer means and recorded in a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, and the like, alone or in combination. The program instructions to be recorded on the medium may be those specially designed and configured for the embodiments or may be available to those skilled in the art of computer software. Examples of computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape; optical media such as CD-ROMs and DVDs; magnetic media such as floppy disks; Includes hardware devices that are specially configured to store and execute program instructions such as magneto-optical media and ROM, RAM, flash storage, and the like. Examples of program instructions include machine language code such as those produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter or the like. The hardware devices described above may be configured to operate as one or more software modules to perform the operations of the embodiments, and vice versa.

Although the embodiments have been described with reference to the drawings, various technical modifications and variations may be applied to those skilled in the art. For example, it is to be understood that the techniques described may be performed in a different order than the described methods, and / or that components of the described systems, structures, devices, circuits, Lt; / RTI > or equivalents, even if it is replaced or replaced.

Claims (20)

A feature extraction unit for extracting a nonlinear feature of an image;
An update unit for performing a repetitive update on the form factor based on the regression factor matrix obtained by training and the nonlinear feature; And
A position detector for detecting the feature point position of the image based on the updated shape factor and the statistical shape model obtained by training;
And a feature point positioning device. The method according to claim 1,
Wherein,
Acquiring a similarity transformation matrix by aligning the intermediate shape and the average shape acquired through the update, and performing a similarity transformation on the intermediate shape and the detected image based on the similarity transformation matrix
Feature point positioning device. 3. The method of claim 2,
Wherein,
Resolution polyramide frame, and performing a minutia positioning from the image of the first resolution to a threshold, and aligning the result of positioning the minutiae to a second resolution larger than the first resolution And the similarity transformation matrix is obtained by positioning the minutiae points
Containing
Feature point positioning device. The method of claim 3,
Wherein,
The feature points acquired on the basis of the similarity transformation matrix obtained through the iterative updating are inversely transformed into the coordinate system of the detected image
Feature point positioning device. 3. The method of claim 2,
Wherein,
The similarity transformation matrix used in the iterative update of the k-th type coefficient

Target function

Lt; / RTI >

Is the i-th sample form after the k-th iteration
Feature point positioning device. The method according to claim 1,
Wherein,
Determining a regression factor matrix based on the optimal shape coefficient update amount of the training sample and the nonlinear feature
Feature point positioning device. The method according to claim 6,
Wherein the nonlinear feature comprises a nonlinear texture feature;
Determining a nonlinear characteristic of the training sample comprises:
Learning an average texture feature in the training sample;
Determining a difference value between the nonlinear texture feature extracted from the training sample and the average texture feature as nonlinear characteristics of the training sample
Wherein the feature point positioning device comprises: The method according to claim 6,
Wherein the nonlinear texture feature comprises a type of nonlinear texture feature or a combination of at least two types of nonlinear texture features;
Using a combination of the same or different nonlinear texture features or at least two nonlinear texture features at different iteration stages
Feature point positioning device. The method according to claim 6,
Wherein,
The error between the optimal shape coefficient update amount and the shape coefficient update amount calculated based on the nonlinear characteristic is minimized to determine the regression factor matrix
Feature point positioning device. 10. The method of claim 9,
The update unit
The regression factor matrix used for iterative update of the kth form factor

Target function

Lt; / RTI >
Here, N means the total number of training samples,

Is the optimal shape coefficient update amount of the i < th > sample at the k <

Is a non-linear characteristic of the i-th sample at the k-th iteration,

Is a form factor update amount obtained by calculating based on the nonlinear characteristic of the i < th > sample at the k <
Feature point positioning device. Extracting a nonlinear characteristic of the image;
Repeatedly updating the form factor based on the regression factor matrix obtained through training and the non-linear feature;
Detecting the feature point of the image based on the updated shape coefficient and the statistical shape model obtained by training
/ RTI > 12. The method of claim 11,
The step of repeatedly updating the form factor comprises:
Obtaining an affinity transformation matrix by arranging an intermediate form and an average form obtained through an update;
Performing a similarity transformation on the intermediate form and the detected image based on the similarity transformation matrix
Further comprising
How to position the minutiae. 13. The method of claim 12,
Wherein the obtaining the similarity transformation matrix comprises:
Aligning an intermediate shape and an average shape acquired through an update using a multiresolution pyramid frame; And
Positioning the minutiae point in the image of the first resolution to a threshold value and mapping the result of the positioning of the minutiae to the image of the second resolution larger than the first resolution,
Containing
How to position the minutiae. 14. The method of claim 13,
Transforming the feature points acquired based on the similarity transformation matrix obtained through the iterative update to the coordinate system of the detected image
Further comprising
How to position the minutiae. 13. The method of claim 12,
The similarity transformation matrix used in the iterative update of the k-th type coefficient

Target function

Lt; / RTI >

Is the i-th sample form after the k-th iteration
How to position the minutiae. 12. The method of claim 11,
Establishing a nonlinear characteristic of the training sample; And
Obtaining a regression factor matrix based on an optimal shape factor update amount of the training sample and the nonlinear feature;
How to position the minutiae. 17. The method of claim 16,
Wherein the nonlinear feature comprises a nonlinear texture feature;
Determining a nonlinear characteristic of the training sample comprises:
Learning an average texture feature in the training sample;
Determining a difference value between the nonlinear texture feature extracted from the training sample and the average nonlinear texture feature as nonlinear characteristics of the training sample
/ RTI > 17. The method of claim 16,
Wherein the nonlinear texture feature comprises a combination of one or more types of nonlinear texture features;
Using the combination of the one or more non-linear texture features when repeating the update
How to position the minutiae. 17. The method of claim 16,
Determining a regression factor matrix based on an optimal shape coefficient update amount of the training sample and the non-
Determining a regression factor matrix by minimizing an error between an optimal shape factor update amount and a form factor update amount calculated based on the nonlinear feature;
Containing
How to position the minutiae. 20. The method of claim 19,
The regression factor matrix used for iterative update of the kth form factor

Target function

Lt; / RTI >
N is the total number of training samples,

Is the optimal shape coefficient update amount of the i < th > sample at the k <

Is a non-linear characteristic of the i-th sample at the k-th iteration,

Is a form factor update amount obtained by calculating based on the nonlinear characteristic of the i < th > sample at the k <
How to position the minutiae.

Images