KR101914244B1 - Apparatus and method for image segmentation using dual dictionary learning - Google Patents

Abstract

An image segmentation method according to an embodiment of the present invention includes: acquiring a segmentation target image to be divided into a plurality of segments; acquiring a first dictionary, which is a set of one or more filters indicating characteristics of the sample image; Acquiring a second dictionary, which is a set of one or more filters indicating characteristics of a divided sample image generated by dividing the sample image into a plurality of parts, using the first dictionary to generate a sparse code generating a resulting sparse code set that is a set of sparse codes, and generating a resultant image by partitioning the divided object image into a plurality of portions using the second dictionary and the resulting sparse code set have.

Description

[0001] APPARATUS AND METHOD FOR IMAGE SEGMENTATION [0002] USING DUAL DICTIONARY LEARNING [0003]

The present invention relates to a bright-field microscope image in which a plurality of objects such as a human cell are photographed, wherein the image is divided into a plurality of portions according to an area occupied by each of the plurality of objects in the image The present invention relates to an apparatus and a method for quickly and accurately performing a process for acquiring an image.

Cell segmentation is the process of accurately segmenting the area occupied by each cell in a cell image taken using an optical device such as a bright field microscope. Cell division is a very important part of the medical image processing process because the diagnosis and progress of the disease can be confirmed by performing the cell division operation. Originally, such cell division work has been carried out manually by a skilled expert, but such manual work has problems in that accuracy and accuracy are guaranteed, while consuming a lot of time and effort.

Therefore, studies have been actively carried out to automate cell division. As a result, techniques such as a level set model algorithm and a Chan-Vese model algorithm have been devised. However, these techniques require prior knowledge such as cell edge information or nucleation location, and the accuracy of the technique is likely to be reduced when applied to complex images.

On the other hand, another method of automating the cell division operation is a method in which a machine learning technique such as a deep neural network (DNN) is applied. The use of machine learning techniques does not require prior knowledge of the image, and has the advantage that the more the learning data is accumulated, the higher the accuracy. However, machine learning techniques also have a disadvantage of requiring a high-performance system due to a large amount of learning data and requiring a long time for the operation.

Patent Document: Korean Patent Laid-Open No. 10-2010-0116404 (published on November 11, 2010)

SUMMARY OF THE INVENTION It is an object of the present invention to provide a method and apparatus for performing a segmentation operation on an image of a plurality of objects, in particular, a cell division operation on an image obtained by photographing living cells including a human being, Apparatus and method.

It is to be understood, however, that the scope of the present invention is not limited to the above-mentioned objects, and may be embodied from the following description, which may be understood by those skilled in the art to which the present invention belongs. have.

An image segmentation method according to an embodiment of the present invention includes: acquiring a segmentation target image to be divided into a plurality of segments; acquiring a first dictionary, which is a set of one or more filters indicating characteristics of the sample image; Acquiring a second dictionary, which is a set of one or more filters indicating characteristics of a divided sample image generated by dividing the sample image into a plurality of parts, using the first dictionary to generate a sparse code generating a resulting sparse code set that is a set of sparse codes, and generating a resultant image by partitioning the divided object image into a plurality of portions using the second dictionary and the resulting sparse code set have.

Also, the number of filters included in the first dictionary is K, the first dictionary is obtained based on K sparse codes included in the first sparse code set, and the first dictionary and the first sparse code set Is a result of performing a convolution operation between k (k is an integer of 0 or more and less than K) first filter of the first dictionary and kth rare code of the first sparse code set for each k The difference between the obtained image and the sample image can be minimized.

Also, the number of filters included in the second dictionary is K, and the second dictionary is obtained based on K sparse codes included in the second sparse code set, and the second dictionary and the second sparse code set Is obtained by adding all the results obtained by performing a convolution operation between k (k is an integer of not less than 0 and less than K) filter of the second dictionary and kth rare code of the second sparse code set for each k, and , And the difference between the divided sample images can be minimized.

When the number of filters included in the first dictionary is K, the resultant sparse code set includes K sparse codes, and the resultant sparse code set includes k (k is a value between 0 and K The difference between the image obtained by adding all of the results obtained by performing the convolution operation between the k-th rare code of the resultant sparse code set and the k-th rare code of the resultant sparse code set to k, and the division target image may be minimized .

Also, when the number of filters included in the second dictionary is K, the resultant sparse code set includes K sparse codes, and the resultant image includes k (k is an integer greater than or equal to 0 and less than K K) < th > filter of the resultant sparse code set and the result of performing a convolution operation between k < th >

The method may further comprise the step of replacing the convolution operation with a matrix multiplication operation using a Fourier transform to obtain the result of the convolution operation.

The method may further include the step of receiving an arbitrary natural number and determining the value of K as the same number as the arbitrary natural number.

The division target image and the sample image are images of a plurality of cells, and the division sample image is divided into a plurality of portions based on contours of images of a plurality of cells included in the sample image It can be segmented video.

The resultant image may be a binarized image such that the contour of the image of each of the plurality of cells in the divided object image and the portion of the plurality of cells in the divided object image excluding the contour are distinguished from each other .

An image segmentation apparatus according to an embodiment of the present invention includes an input unit for acquiring a segmentation target image to be divided into a plurality of segments, a first dictionary, which is a set of one or more filters indicating the characteristics of the sample image, Which is a set of one or more filters representing a characteristic of a divided sample image generated by dividing the sample image into a plurality of parts, And a result image generation unit for generating a result image by dividing the division target image into a plurality of parts using the dictionary and the resultant sparse code set.

Further, when the number of filters included in the first dictionary is K, and the first dictionary is obtained based on K sparse codes included in the first sparse code set, the apparatus further comprises: the difference between the image obtained by adding all of the results obtained by performing the convolution operation between k (k is an integer equal to or more than 0 and less than K) filter and the kth rare code of the first sparse code set for k, A database control unit for determining the first dictionary and the first sparse code set to be minimum, and a database for storing the first dictionary and the first sparse code set.

Further, when the number of filters included in the second dictionary is K and the second dictionary is obtained based on K sparse codes included in the second set of rare codes, the apparatus further comprises: a difference between the image obtained by adding all of the results obtained by performing a convolution operation between k (k is an integer equal to or more than 0 and less than K) filter and the kth rare code of the second sparse code set, The second dictionary and the second sparse code set so that the second dictionary and the second sparse code set are minimized, and a database storing the second dictionary and the second sparse code set.

When the number of filters included in the first dictionary is K, the resultant sparse code set includes K sparse codes, and the sparse code calculator calculates k (k is an integer in the range of 0 to K Of the resultant sparse code set and a result of performing a convolution operation between kth sparse code of the resultant sparse code set and a result of the convolution operation for each k, Can be determined.

When the number of the filters included in the second dictionary is K, the resultant sparse code set includes K sparse codes, and the resultant image output unit outputs k And the convolution operation between the kth scarce code of the resultant sparse code set and the kth sparse code of the resulting sparse code set is performed for each k.

Also, the result of the convolution operation may be calculated by replacing the convolution operation with a matrix multiplication operation using Fourier transform.

Also, the input unit receives an arbitrary natural number, and the value of K may be determined to be the same as the arbitrary natural number.

The division target image and the sample image are images obtained by photographing a plurality of cells, and the division sample image may include an image segmented into a plurality of segments based on contours of images of a plurality of cells included in the sample image Lt; / RTI >

In addition, the resultant image may be a binarized image such that a contour of an image of each of a plurality of cells in the divided object image and a portion of the plurality of cells in the divided object image, excluding the contour, are distinguished from each other.

According to the embodiment of the present invention, by using both a dictionary composed of a filter showing the characteristics of a sample image before a segment and two dictionaries made of a filter showing the characteristics of a sample image after segmentation, By introducing a dual dictionary learning technique that performs segmentation, cell segmentation work on cell images can be performed with high accuracy. Further, after building the learned dictionary from the sample images before and after the segment, the segmentation target image can be segmented by a simple operation using only the pre-built dictionary without additional learning, and therefore, the work efficiency can also be improved.

FIG. 1 is a diagram showing a cell image according to a conventional technique and a result of cell division on the image.
FIG. 2 is a conceptual illustration of convolutional sparse coding (CSC) according to an embodiment of the present invention.
FIG. 3 is a diagram for conceptually illustrating a method of dividing an image by a dual prior learning technique according to an embodiment of the present invention.
4 is a diagram illustrating the configuration of an image segmentation apparatus according to an embodiment of the present invention.
5 is a diagram illustrating a procedure of an image segmentation method according to an embodiment of the present invention.
FIG. 6 is a diagram for conceptually explaining a process for obtaining a preliminary variable G in an image segmentation method according to an embodiment of the present invention.
7 is a view for explaining a post-processing technique of the image segmentation method according to an embodiment of the present invention.
FIG. 8 is a view for explaining a result of performing an image segmentation method according to an embodiment of the present invention.

BRIEF DESCRIPTION OF THE DRAWINGS The advantages and features of the present invention, and the manner of achieving them, will be apparent from and elucidated with reference to the embodiments described hereinafter in conjunction with the accompanying drawings. The present invention may, however, be embodied in many different forms and should not be construed as 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. To fully disclose the scope of the invention to those skilled in the art, and the invention is only defined by the scope of the claims.

In the following description of the present invention, a detailed description of known functions and configurations incorporated herein will be omitted when it may make the subject matter of the present invention rather unclear. The following terms are defined in consideration of the functions in the embodiments of the present invention, which may vary depending on the intention of the user, the intention or the custom of the operator. Therefore, the definition should be based on the contents throughout this specification.

FIG. 1 is a diagram showing a cell image according to a conventional technique and a result of cell division on the image. 1 shows an image representing a plurality of cells obtained through a bright field microscope, which can be a segmentation image 11 that has not yet undergone cell division. The right figure of FIG. 1 is a result image 12 obtained by performing cell division on the division target image 11, and is a result obtained through the above-described level set model algorithm.

In the result image 12, it can be seen that the result of cell division is shown by dividing the area occupied by the plurality of cells in the image by using the closed curve 13. That is, each closed curve 13 represents the outline of each cell in the image. However, according to the conventional techniques including the level set model algorithm, the quality or the accuracy of the resulting image 12 is not guaranteed.

Therefore, an image segmentation method based on the dual prior learning technique according to an embodiment of the present invention aims at obtaining the result of image segmentation corresponding to the result image 12 accurately and quickly as compared with the conventional method. However, the result image 12 of FIG. 1 expresses the result of cell segmentation using the closed curve 13 representing the outline of each cell, but in the case of the image segmentation method according to an embodiment of the present invention, no.

FIG. 2 is a conceptual illustration of convolutional sparse coding (CSC) according to an embodiment of the present invention. Referring to FIG. 3, a redundant learning method according to an embodiment of the present invention is described. Referring to FIG. 2, a filter representing a feature of an image and a rare- Let's take a look at coding.

2, the original image 21 may have a predetermined resolution (here, 256 pixels in the horizontal direction and 256 pixels in the vertical direction) represented by the number of pixels. Also, assuming that the original image 21 is a monochrome image, the value (brightness) of each pixel of the original image 21 can be expressed by an integer value. For example, the value of each pixel can be represented by one of 256 values from 0 (black) to 255 (white). Such an original image 21 can be approximated by using K (K is a natural number) filters d and K sparse codes (x).

Each filter d is an image representing a feature of the original image 21 and may include a smaller number of pixels (here, 11 pixels in the horizontal direction and 11 pixels in the vertical direction) in comparison with the original image 21 . A set of such filters can be defined as a dictionary (31). That is, the dictionary 31 can be regarded as a set of information about the original image 21.

The sparse code (x) can also be regarded as an image, which may generally include a number of pixels equal to or similar to the original image 21 (256 pixels in this case and 256 pixels in height). In the sparse code (x), the value of most pixels is zero, and only a very small number of pixels have non-zero values. That is, according to the sparse code (x), all parts except the specific part of the image are zero, and it is possible to extract the characteristic of the specific part of the image effectively by using the sparse code (x). It also has the advantage that the amount of computation of the sparse code (x) can be reduced.

As described above, when the number of the filter (d) and the number of the sparse codes (x) are K, the product product of the kth (k is an integer of not less than 0 and less than K) filter d and the kth rare code (K) so that the approximated image obtained by adding all the results of performing the convolution operation to each k is as close as possible to the original image 21, that is, the difference between the approximated image and the original image 21 is minimized. (D) and the sparse code (x) can be defined. Here, the difference between images may be a value obtained by adding, for example, a difference between values of pixels having the same positions in two images to be searched for all pixels.

Thus, each of the K filters d and the sparse code x that can best represent the original image 21 becomes a filter d and a sparse code x that satisfy the following expression (1). For reference, the value of K may be determined based on the input from the user or the performance of the system in which the image segmentation apparatus 100 is implemented.

In Equation (1), S represents the original image 21, d _k and x _k represent the k-th filter (d) and the k-th rare code (x), respectively. It is an optimized value. That is, obtaining the filter (d) and the sparse code (x) as a solution satisfying the expression (1) has substantially the same meaning as that of obtaining an approximate image closest to the original image (21).

) 역시 만족하여야 한다는 뜻이다. 이러한 기호 및 인자에 관한 설명은 아래의 다른 수학식들에도 유사하게 적용될 수 있다.For reference, "min" of symbols and is the factor of "d, x" below, min sign to the value of the filter (d) with rare code (x) such that expression is at least on the right side (more specifically, each d _k And the value of x _k ). The term " st "is an abbreviation of " such that "

) Must also be satisfied. The description of these symbols and factors can be applied similarly to the other mathematical formulas below.

FIG. 3 is a diagram for conceptually illustrating a method of dividing an image by a dual prior learning technique according to an embodiment of the present invention. The sample image 210 of FIG. 3 may be an image that has not yet been segmented, for example, a cell image obtained through a bright field microscope. The divided sample image 220 is an image generated by dividing the sample image 210 into a plurality of portions. When the sample image 210 is a cell image, the divided sample image 220 is manually generated, It can be an image expressed.

That is, the sample image 210 and the divided sample image 220 may correspond to the division target image 11 and the result image 12, respectively, in FIG. Although there may be a plurality of the sample images 210 and the divided sample images 220, the following description will be made assuming that there is one for each of the sample images 210 and the divided sample images 220 for convenience of explanation.

The sample image 210 includes a first dictionary 310 which is a set of filters D representing characteristics of the sample image 210 and a first sparse code set 410 which is a set of the corresponding rare codes X ₁ Can be approximated. 3, it is assumed that the number of the filters D included in the first dictionary 310 and the number of the rare codes X ₁ included in the first sparse code set 410 are K An approximate image obtained by adding all the results of performing a convolution operation between k (k is an integer equal to or more than 0 and less than K) filter D and a kth rare code X ₁ for each k is obtained as a sample image 210 (D) and the sparse code (X ₁ ) so that the difference between the approximated image and the sample image 210 is minimized.

Similarly, the divided sample image 220 includes a second dictionary 320, which is a set of filters B showing the characteristics of the divided sample image 220, and a second rare (which is a set of the corresponding rare code X ₂ ) Can be approximated using a set of codes 420. The number of the filters B included in the second dictionary 320 and the number of the rare codes X ₂ included in the second rare code set 420 are respectively K . In this case, the approximate image obtained by adding the results of performing the convolution operation between k (k is an integer equal to or more than 0 and less than K) filter (B) and the kth rare code (X ₂ ) (B) and the sparse code (X ₂ ) so that the difference between the approximated image and the divided sample image 220 is minimized.

The filters D and B and the sparse codes X ₁ and X _{2 that} best represent the sample image 210 and the divided sample image 220 are determined to satisfy the following Equation 2 .

In Equation (2), S represents a sample image 210 and L represents a divided sample image 220, and α, γ, λ ₁ , λ ₂ and τ are arbitrary constants and optimized through experiments. For the sake of convenience, the sigma (Σ) symbols and the index k are omitted for D, B, X ₁ , and X ₂ .

(D 1, B) and the sparse code (X ₁ , X ₂ ) as a solution satisfying the expression ( ₂ ) can be obtained by obtaining an approximate optimal image for each of the sample image 210 and the divided sample image 220 It has virtually the same meaning as. In order to solve Equation (2), the values of the filters (D, B) may be randomly set and then the values of the filters (D, B) satisfying the Equation (2) Will be described in detail later.

4 is a diagram illustrating the configuration of an image segmentation apparatus according to an embodiment of the present invention. 4 may include an input unit 110, a database 120, a database control unit 130, a sparse code calculation unit 140, and a resultant image generation unit 150. However, since the image dividing apparatus 100 of FIG. 4 is only one embodiment of the present invention, the concept of the present invention is not limited to FIG.

The input unit 110 may receive the partitioning target image 230 to be divided into a plurality of portions. The input unit 110 may be implemented by various devices, and may include a data bus or a communication module for receiving input data including a division target image 230 from outside the image dividing device 100 , An imaging device such as a scanner or a camera for physically acquiring or directly generating the division target image 230, or a bright field microscope. The sample image 210 and the divided sample image 220 may also be acquired by the input unit 110 and stored in a database 120 to be described later.

The input unit 110 may store the number of the filters D and B included in the first dictionary 310 and the second dictionary 320 or the first sparse code set 410 and the second sparse code set 420, The number of the rare codes (X ₁ , X ₂ ) included in each of them can be inputted from the user of the image dividing device 100. Accordingly, the input unit 110 may include an input device such as a keyboard, a mouse, a touch screen, and a touch pad.

The database 120 may store a sample image 210 and a split sample image 220. The database 120 also includes a first dictionary 310, a second dictionary 320, a first sparse code set 410, and a second sparse code, which are data based on the sample image 210 and the split sample image 220. [ Set 420, and so on. The database 120 may be embodied as a computer-readable recording medium. Examples of such computer-readable media include magnetic media such as hard disks, floppy disks and magnetic tape, optical media such as CD-ROMs, DVDs, A magneto-optical media such as a floppy disk, a magneto-optical media such as a floppy disk, a magneto-optical media such as a floppy disk, a magneto-optical media such as a floppy disk, a magneto-optical media such as a floppy disk, a flash memory, .

The database control unit 130 extracts the first dictionary 310 and the second dictionary 310 from the sample image 210 and the divided sample image 220 stored in the database 120 based on the above- 320, a first sparse code set 410, a second sparse code set 420, and the like. The data determined by the database control unit 130 may be stored in the database 120. [ The database control unit 130 may be implemented by a computing device including a microprocessor due to the nature of its operation.

On the other hand, the image segmentation apparatus 100 may not include the database 120 and the database control unit 130. In this case, the image segmentation apparatus 100 may include a first dictionary 310, a second dictionary 320, a first sparse code set 410, and a second sparse code set 420, And can be utilized from outside.

The sparse code calculation unit 140 can generate the resulting sparse code set 430 that is a set of sparse codes for the division target image 230 using the first dictionary 310. [ The resultant image generating unit 150 may generate the resultant image 240 in which the divided object image 230 is divided into a plurality of parts by using the second dictionary 320 and the resultant sparse code set 430 have. That is, the sparse code calculation unit 140 and the resultant image generation unit 150 may perform the segmentation operation on the segmentation target image 230 (the division target image 230 is the cell image segment) using the data constructed in the database 120 Cell division work) can be performed. The sparse code calculation unit 140 and the resultant image generation unit 150 may be implemented by a computing device including a microprocessor as well as the database control unit 130, and detailed operations thereof will be described later.

5 is a diagram illustrating a procedure of an image segmentation method according to an embodiment of the present invention. The image dividing method of FIG. 5 may be performed by the image dividing apparatus 100 described with reference to FIG. However, since the method shown in FIG. 5 is only an embodiment of the present invention, the concept of the present invention is not limited to FIG. 5, and each step of the method shown in FIG. And may be performed in a different order of the bar and the order.

First, the input unit 110 can acquire the division target image 230 (S110). The database controller 130 may determine the first dictionary 310 and the first sparse code set 410 from the sample image 210 stored in the database 120 at step S120. In addition, the database controller 130 may determine the second dictionary 320 and the second sparse code set 420 from the split sample image 220 stored in the database 120 (S130).

The above steps S120 and S130 are substantially the same as those for obtaining the solution of the above-mentioned equation (2). Hereinafter, the process of solving Equation (2) will be described in more detail.

In order to divide equation (2) into sub-problems, a preliminary variable Y (Y ₁ , Y ₂ ) and a double variable U (U ₁ , U ₂ ) may be introduced, 3, and ρ ₁ and ρ ₂ are values optimized through experimentation with arbitrary constants.

Equation (3) can be divided into independent details according to each variable. In order to obtain X ₁ , it is possible to extract the term related to X ₁ from Equation (3) to derive Equation (4) as follows.

In order to facilitate the calculation of Equation (4), it is possible to transform all variables into frequency domain variables using a Fourier transform so that the convolution operation can be converted to a simpler matrix multiplication operation have. Equation (5) below is the result of transforming Equation (4) into Equation in the frequency domain.

If the differential operation for X _f1 is applied to Equation (5), Equation (6) can be derived as follows.

In the above equation (6), D _f = [diag (d _f1 ), ..., diag (d _fk )] and D _f ^H is a complex conjugated transpose of D _f . X _f1, which satisfies Equation 6, can be easily calculated using the Sherman-Morrison formula, which yields X ₁ .

Also, in order to obtain Y ₁ in Equation (3), it is possible to derive the following Equation (7) by extracting the term relating to Y ₁ from Equation (3).

In Equation (7), the value of Y ₁ can be obtained by the following Equation (8) using a shrinkage operation.

In order to obtain the double variable U ₁ , it is also possible to derive the following Equation (9) by extracting the term relating to U ₁ from Equation (3).

U ₁ obtained from Equation (9) can be updated as shown in Equation (10) based on X ₁ and Y ₁ obtained as above, and such updating is performed when the value of X ₁ -Y ₁ is below a certain level To converge to.

For reference, as the expression obtained by adding the value of X ₁ -Y ₁ of the value of the existing value of U ₁ it is the means to update the value of U _1. That is, the equal sign (=) in Equation (10) functions as an assignment operator, not an operator in the sense that the equal sign is the same in both equations.

The first sparse code set 410 can be determined by obtaining X ₁ satisfying the expression (2) through a series of the processes described so far. Hereinafter, by obtaining D satisfying the expression (2) 310 will be described. To do this, first, a preliminary variable G and a double variable H are introduced into the equation (2), and a port relating to D is extracted to derive the following equation (11), similarly to the formula .

Equation (11) can also be transformed into Equation (12) to be expressed in the frequency domain using the Fourier transform, for simplification of the convolution operation.

When the differential operation for D _f is applied to Equation (12), the following Equation (13) is derived, from which the value of D _f can be obtained.

Similar to the case of Equation 6, X _f1 = [diag (x _f1 ), ..., diag (x _fk )] and X _f1 ^H is the complex conjugate transpose of X _f1 . D _f1, which satisfies Equation 13, can also be calculated through the Sherman-Morrison formula.

In order to obtain the preliminary variable G, Equation (14) below can be solved.

FIG. 6 is a diagram for conceptually explaining a process for obtaining a preliminary variable G in an image segmentation method according to an embodiment of the present invention. G obtains the G to obtain a D by performing an inverse Fourier transform (inverse Fourier transform) on D _f found by the equation (13), cut to the size D of the filters defined in advance as shown in Fig. The G thus obtained can be expanded to the same size as that of the sample image 210, and the extension can be performed by filling the portion other than G obtained by the trimming with 0s. Finally, a normalization process may be performed on the extended G, and the normalized G may be determined as a final G value.

On the other hand, Equation 15 can be solved to obtain the double variable H as follows.

Similar to the case of U ₁ described above, H obtained through Equation (15) can be updated as shown in Equation (16) below based on D and G obtained above, It can be continued until it converges below the level.

The first dictionary 310 and the first sparse code set 410 can be determined by obtaining X ₁ and D through the process described so far and the second dictionary 320 and the second sparse code set 420 can also be determined It can be confirmed through a similar process.

When the steps S120 and S130 are completed, the sparse code calculation unit 140 can obtain the resultant sparse code set 430 using the first dictionary 310 and the second dictionary 320 (S140). The resulting sparse code set 430 includes the same number of resultant sparse codes X as the number K of filters D and B each of which includes the first dictionary 310 and the second dictionary 320 And the resulting sparse code set 430 can be computed by finding the K resultant sparse codes X. [ The calculation of the resulting sparse code set 430 can be performed by the following Equation (17), and the index k is omitted for the sake of convenience in Equation (17).

Finally, the resultant image 240 can be generated using Equation (18) using the calculated result set of scarce code 430 (S150). L in Equation (18) represents the resulting image 240. [

That is, by learning the existing sample image 210 and the divided sample image 220 as an accurate division result for the sample image 210, the first dictionary 310 and the second dictionary The first dictionary 310 and the second dictionary 320 obtained as described above can be used to perform a complicated learning process on the new division target image 230 without repeating the learning process of Equation 17 And the resultant image 240 can be easily obtained by a simple convolution operation such as < RTI ID = 0.0 > 18. ≪ / RTI > Also, as mentioned above, the number of the sample images 210 can be plural, and the reliability of the first dictionary 310 and the second dictionary 320 can be further improved by using a sufficient number of the sample images 210 Therefore, it is possible to obtain a more accurate result image 240.

7 is a view for explaining a post-processing technique of the image segmentation method according to an embodiment of the present invention. Figs. 7A and 7B show an example division target image 230 and a result image 240, respectively, each of which is an image of a cell. The resulting image 240 may be edited into a binarized image as shown in FIG. 7F.

A brief description of the binarization process as the post-process is as follows. 7 (C), only the portion corresponding to the cell membrane (the contour portion of each cell in the image) is extracted from the resultant image 240 of FIG. 7 (B) by using the multistep threshold value based on the Otsu threshold value. As shown in Fig. Next, in order to connect the parts of the cell membrane that are not completely connected to each other, a morphology operation such as erosion, dilation, open, D).

In addition, to extract only a thin cell membrane region, the cell membrane expressed in thick by using the above-mentioned morphology operation can be represented as thin as shown in (E) of FIG. Finally, the final image binarized by using the logical operation on the image of (E) of FIG. 7 can be obtained as shown in (F) of FIG. That is, according to the finalized image of the binarized image, the outline of each cell and the portion excluding the outline in each cell are distinguished from each other, so that the area occupied by each cell can be clearly recognized.

FIG. 8 is a view for explaining a result of performing an image segmentation method according to an embodiment of the present invention. 8 (A) and 8 (D) are luminescence microscope images obtained by culturing a human brain tumor cell in an antibody and culturing the same, and FIGS. 8 (B) and 8 (C) and (F) are the result of cell division by DNN (deep neural network), which is one of the conventional deep-running methods.

In addition, (B) and (C) are the results for (A), and (E) and (F) are for (D). The results were derived from MATLAB on systems with NVIDIA GTX Titan X (12GB) and Linux 14.04. Also, the first dictionary 310 and the second dictionary 320 were set to include 225 filters D and B, respectively, and the size of each filter was set to 11 pixels horizontally and 11 pixels vertically.

Referring to FIG. 8, the cell division result obtained by the image segmentation method according to an embodiment of the present invention shows a region occupied by each cell more finely (i.e., thinner cell membrane) than the result of cell division by DNN can see. In addition, while it took 1.5 days under the above conditions to derive the result by the DNN-based cell division method, the result of the image segmentation method according to the embodiment of the present invention is 58 minutes when the CPU is used and the GPU is used The case took 12 minutes. That is, according to the image segmentation method according to an embodiment of the present invention, it is possible to perform cell segmentation more quickly and accurately than the conventional method.

Combinations of each step of the flowchart and each block of the block diagrams appended to the present invention may be performed by computer program instructions. These computer program instructions may be loaded into a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus so that the instructions, which may be executed by a processor of a computer or other programmable data processing apparatus, And means for performing the functions described in each step are created. These computer program instructions may also be stored in a computer usable or computer readable memory capable of directing a computer or other programmable data processing apparatus to implement the functionality in a particular manner so that the computer usable or computer readable memory It is also possible for the instructions stored in the block diagram to produce a manufacturing item containing instruction means for performing the functions described in each block or flowchart of the block diagram. Computer program instructions may also be stored on a computer or other programmable data processing equipment so that a series of operating steps may be performed on a computer or other programmable data processing equipment to create a computer- It is also possible that the instructions that perform the processing equipment provide the steps for executing the functions described in each block of the block diagram and at each step of the flowchart.

Also, each block or each step may represent a module, segment, or portion of code that includes one or more executable instructions for executing the specified logical function (s). It should also be noted that in some alternative embodiments, the functions mentioned in the blocks or steps may occur out of order. For example, two blocks or steps shown in succession may in fact be performed substantially concurrently, or the blocks or steps may sometimes be performed in reverse order according to the corresponding function.

The above description is merely illustrative of the technical idea of the present invention, and various modifications and changes may be made by those skilled in the art without departing from the essential characteristics of the present invention. Therefore, the embodiments disclosed in the present invention are intended to illustrate rather than limit the scope of the present invention, and the scope of the technical idea of the present invention is not limited by these embodiments. The scope of protection of the present invention should be construed according to the following claims, and all technical ideas within the scope of equivalents thereof should be construed as falling within the scope of the present invention.

100: image dividing device
110: input unit
120: Database
130:
140: Sparse code calculation unit
150:

Claims (20)

Obtaining a division target image to be divided into a plurality of portions;
Obtaining a first dictionary that is a set of one or more filters that characterize a sample image;
Obtaining a second dictionary, which is a set of one or more filters that characterize the divided sample images generated by dividing the sample image into a plurality of portions;
Generating a resultant sparse code set that is a set of sparse codes for the segmentation target image using the first dictionary; And
Generating a resultant image by partitioning the segmentation target image into a plurality of segments using the second dictionary and the resultant sparse code set,
Wherein the number of filters included in the first dictionary is K, the first dictionary is obtained based on K sparse codes included in the first sparse code set,
Wherein the first dictionary and the first sparse code set comprise:
A convolution operation between k (k is an integer of 0 or more and less than K) first filter of the first dictionary and kth rare code of the first sparse code set is performed for each k, And the difference between the sample images is set to be minimum
Image segmentation method.
delete Obtaining a division target image to be divided into a plurality of portions;
Obtaining a first dictionary that is a set of one or more filters that characterize a sample image;
Obtaining a second dictionary, which is a set of one or more filters that characterize the divided sample images generated by dividing the sample image into a plurality of portions;
Generating a resultant sparse code set that is a set of sparse codes for the segmentation target image using the first dictionary; And
Generating a resultant image by partitioning the segmentation target image into a plurality of segments using the second dictionary and the resultant sparse code set,
The number of filters included in the second dictionary is K and the second dictionary is obtained based on K sparse codes included in the second set of rare codes,
The second dictionary and the second set of rare codes,
A convolution operation between k (k is an integer of 0 or more and less than K) second filter of the second dictionary and kth rare code of the second sparse code set is performed for each k, And the difference between the sample images is set to be minimum
Image segmentation method. Obtaining a division target image to be divided into a plurality of portions;
Obtaining a first dictionary that is a set of one or more filters that characterize a sample image;
Obtaining a second dictionary, which is a set of one or more filters that characterize the divided sample images generated by dividing the sample image into a plurality of portions;
Generating a resultant sparse code set that is a set of sparse codes for the segmentation target image using the first dictionary; And
Generating a resultant image by partitioning the segmentation target image into a plurality of segments using the second dictionary and the resultant sparse code set,
When the number of filters included in the first dictionary is K, the resulting sparse code set includes K sparse codes,
The resultant sparse code set is obtained by performing a convolution operation between k (k is an integer of 0 or more and less than K) th filter of the first dictionary and the kth rare code of the resulting sparse code set for each k And the difference between the obtained image and the division target image is set to be minimum
Image segmentation method. Obtaining a division target image to be divided into a plurality of portions;
Obtaining a first dictionary that is a set of one or more filters that characterize a sample image;
Obtaining a second dictionary, which is a set of one or more filters that characterize the divided sample images generated by dividing the sample image into a plurality of portions;
Generating a resultant sparse code set that is a set of sparse codes for the segmentation target image using the first dictionary; And
Generating a resultant image by partitioning the segmentation target image into a plurality of segments using the second dictionary and the resultant sparse code set,
When the number of filters included in the second dictionary is K, the resulting sparse code set includes K sparse codes,
The resultant image is obtained by adding all the results obtained by performing a convolution operation between the kth (k is an integer of 0 or more and less than K) filter of the second dictionary and the kth rare code of the resulting sparse code set for each k Video-in
Image segmentation method. The method according to claim 1,
Further comprising the step of replacing the convolution operation with a matrix multiplication operation using a Fourier transform to obtain the result of the convolution operation
Image segmentation method. The method according to claim 1,
Receiving an arbitrary natural number; And
Further comprising the step of determining the value of K as a number equal to the arbitrary natural number
Image segmentation method. The method according to claim 1,
Wherein the division target image and the sample image are images obtained by photographing a plurality of cells and the division sample image is divided into a plurality of portions based on contours of images of a plurality of cells included in the sample image Video-in
Image segmentation method. 9. The method of claim 8,
Wherein the resultant image is a binarized image such that a contour of an image of each of a plurality of cells in the divided object image and a portion of the plurality of cells in the divided object image excluding the contour are distinguished from each other
Image segmentation method. An input unit for acquiring a division target image to be divided into a plurality of portions;
A sparse code calculation unit for generating a resultant sparse code set which is a set of sparse codes for the segmentation target image, using a first dictionary that is a set of one or more filters indicating characteristics of a sample image;
A second dictionary, which is a set of one or more filters showing the characteristics of the divided sample images generated by dividing the sample image into a plurality of parts, and a second dictionary that is a function of dividing the divided object image into a plurality of parts A result image generating unit for generating a result image;
Wherein when the number of filters included in the first dictionary is K and the first dictionary is obtained based on K sparse codes included in the first sparse code set, k of the first dictionary K) of the first sparse code set and a kth sparse code of the first sparse code set are multiplied by k, A database control unit for determining a dictionary and the first sparse code set; And
And a database storing the first dictionary and the first sparse code set
Image segmentation device. delete An input unit for acquiring a division target image to be divided into a plurality of portions;
A sparse code calculation unit for generating a resultant sparse code set which is a set of sparse codes for the segmentation target image, using a first dictionary that is a set of one or more filters indicating characteristics of a sample image;
A second dictionary, which is a set of one or more filters showing the characteristics of the divided sample images generated by dividing the sample image into a plurality of parts, and a second dictionary that is a function of dividing the divided object image into a plurality of parts A result image generating unit for generating a result image;
Wherein when the number of filters included in the second dictionary is K and the second dictionary is obtained based on K sparse codes included in the second sparse code set, k of the second dictionary K) and the kth scarce code of the second sparse code set are all added to the k-th sparse code and the kth sparse code of the second sparse code set, 2 dictionary and the second sparse code set; And
And a database for storing the second dictionary and the second set of rare codes
Image segmentation device. An input unit for acquiring a division target image to be divided into a plurality of portions;
A sparse code calculation unit for generating a resultant sparse code set which is a set of sparse codes for the segmentation target image, using a first dictionary that is a set of one or more filters indicating characteristics of a sample image; And
A second dictionary, which is a set of one or more filters showing the characteristics of the divided sample images generated by dividing the sample image into a plurality of parts, and a second dictionary that is a function of dividing the divided object image into a plurality of parts And a result image generating unit for generating a result image,
When the number of filters included in the first dictionary is K, the resulting sparse code set includes K sparse codes,
The sparse code calculation unit adds all the results obtained by performing the convolution operation between the kth (k is an integer of 0 or more and less than K) filter of the first dictionary and the kth rare code of the resultant sparse code set for each k The resultant sparse code set is determined such that the difference between the obtained image and the division target image is minimized
Image segmentation device. An input unit for acquiring a division target image to be divided into a plurality of portions;
A sparse code calculation unit for generating a resultant sparse code set which is a set of sparse codes for the segmentation target image, using a first dictionary that is a set of one or more filters indicating characteristics of a sample image; And
A second dictionary, which is a set of one or more filters showing the characteristics of the divided sample images generated by dividing the sample image into a plurality of parts, and a second dictionary that is a function of dividing the divided object image into a plurality of parts And a result image generating unit for generating a result image,
When the number of filters included in the second dictionary is K, the resulting sparse code set includes K sparse codes,
The resultant image output unit adds all the results obtained by performing a convolution operation between the kth (k is an integer of 0 or more and less than K) filter of the second dictionary and the kth rare code of the resultant sparse code set for each k And determines the obtained image as the resultant image
Image segmentation device. 11. The method of claim 10,
The result of the convolution operation is calculated by replacing the convolution operation with a matrix multiplication operation using Fourier transform
Image segmentation device. 11. The method of claim 10,
The input unit receives an arbitrary natural number,
The value of K is determined by the same number as the arbitrary natural number
Image segmentation device. 11. The method of claim 10,
Wherein the division target image and the sample image are images obtained by photographing a plurality of cells and the division sample image is a video image segmented into a plurality of parts based on contours of images of a plurality of cells included in the sample image
Image segmentation device. 18. The method of claim 17,
Wherein the resultant image is a binarized image such that a contour of an image of each of a plurality of cells in the divided object image and a portion of the plurality of cells in the divided object image excluding the contour are distinguished from each other
Image segmentation device. 12. An application program stored on a computer-readable recording medium for performing the respective steps according to the method of claim 1. A computer-readable recording medium on which a program is recorded, the program comprising instructions for performing the respective steps according to the method of claim 1

Images