KR102234364B1 - Storage of deep learning learning images and processing method - Google Patents

Abstract

The present invention discloses a deep learning learning image storage and processing method. A deep learning learning image storage and processing method according to an embodiment of the present invention includes: a first step of capturing and storing at least one image; A second step of extracting a representative value from the stored image; And a third step of generating a learning image by arranging the extracted representative values.

Description

Storage of deep learning learning images and processing method}

The present invention relates to a method for storing and processing a deep learning learning image.

The artificial intelligence (AI) system is a computer system that implements human-level intelligence, and unlike the existing rule-based smart system, the machine learns and judges itself and becomes smarter. As such an artificial intelligence system is used, the recognition rate improves and users' tastes can be more accurately understood, and the existing rule-based smart system is being replaced by an artificial intelligence system based on deep learning for tanks.

Artificial intelligence technology consists of deep learning (machine learning or deep learning) and elemental technologies using deep learning.

Such deep learning (deep learning or machine learning, deep learning) is a high-level abstraction (a task that summarizes features, key contents or functions in a large amount of data or complex data) through a combination of several nonlinear transformation methods. It can be defined as a set of machine learning algorithms that try to do, and is a branch of machine learning that teaches computers how to think. Such deep learning is being used throughout the industry along with artificial intelligence, which means intelligence made from machines.

Meanwhile, as multimedia technology and computer technology develop, users are provided with various services using electronic devices. In particular, as image processing technology gradually develops, users can receive various high-definition images using an image processing device.

However, in the related art, when processing high-quality images through machine learning of an electronic device, the capacity of high-quality images is too large, and unnecessarily large amounts of time and hardware resources have been consumed. In addition, there is a problem that the machine learning speed is significantly lowered compared to the consumption of time and hardware resources.

Korean Patent Publication No. 10-2018-0004898 Korean Patent Publication No. 10-2018-0092494

The present invention was conceived to solve the above problems, and the present invention stores a deep learning learning image capable of improving the deep learning learning speed of a deep learning learning image processing apparatus by compactly reprocessing a large image. And to provide a treatment method for that purpose.

On the other hand, the technical problem to be achieved in the present invention is not limited to the technical problem mentioned above, and other technical problems not mentioned are clearly understood by those of ordinary skill in the technical field to which the present invention belongs from the following description. It will be possible.

As a technical method for achieving the above object, a deep learning learning image storage and processing method according to an embodiment of the present invention includes: a first step of capturing and storing at least one image; A second step of extracting a representative value from the stored image; And a third step of generating a learning image by arranging the extracted representative values.

In one embodiment, the first step is at least one of a microscope image in which the image is collected through a microscope, a fluorescent dyed image dyed with a fluorescent substance through a fluorescent dyeing reagent, and a medical image captured of a medical image of a medical device.

In one embodiment, the second step includes: a 2-1 step of dividing the stored image into at least one image area; And a 2-2 step of extracting a representative value by replacing the image region with a specific coding.

In an embodiment, in step 2-1, the number of bits of the stored image is divided by a uniform number of pixels and divided into image regions.

In one embodiment, the specific coding specifies a representative value for each image area as a 2-digit number or a 4-digit number.

In an embodiment, in step 2-2, ^{the representative value is extracted from the image region through 2 n} , where n is the number of replaced bits.

In one embodiment, step 2-2 is such that adjacent representative values extracted from adjacent image regions do not overlap with each other.

In one embodiment, the specific coding specifies that adjacent representative values extracted from adjacent image regions do not overlap with each other.

In an embodiment, in step 2-2, ^{the representative value is extracted from the image region through 2 n} , where n is the number of bits of the replaced image.

In another embodiment, in the second step, pixels corresponding to the first two digits from the number of pixels of less than half of the bits of the stored image are extracted as a representative value.

In another embodiment, the third step fills or removes an empty space generated when generating a learning image, or adds image information to the empty space.

In another embodiment, the image information is information including the number of fluorescent wavelengths of the image.

According to an embodiment of the present invention, by extracting a representative value from an image, reprocessing the image, and generating the extracted representative value as a deep learning learning image, it is possible to improve a deep learning learning speed for an image.

On the other hand, the effects obtainable in the present invention are not limited to the above-mentioned effects, and other effects not mentioned can be clearly understood by those of ordinary skill in the art from the following description. There will be.

1 is a block diagram showing a schematic configuration of an apparatus for storing and processing a deep learning learning image according to a first embodiment of the present invention.
2 is a block diagram showing a schematic configuration of an apparatus for storing and processing a deep learning learning image according to a second embodiment of the present invention.
3 is a step flow diagram of a method for storing and processing a deep learning learning image according to the first and second embodiments of the present invention.
4 is a detailed step flow diagram of a representative value extraction step according to the first embodiment of the present invention.
FIG. 5 is a diagram illustrating a process of generating an image area in an image area generation step and a representative value extraction process in a representative value extraction step based on a specific coding substitution according to the first embodiment of the present invention.
6 is a detailed step flow diagram of a representative value extraction step according to a second embodiment of the present invention.
7 is a diagram for explaining a process of extracting pixels of an image for each fluorescence wavelength in the pixel extraction step according to the second exemplary embodiment of the present invention.
8 is a diagram for explaining a representative value extraction method in a conditional representative value extraction step and a representative value arranging process in the representative value arranging step according to the second embodiment of the present invention.
9 is a diagram for explaining a representative value including an empty space generated through a representative value arrangement step according to a second exemplary embodiment of the present invention.

Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. DETAILED DESCRIPTION OF THE INVENTION The detailed description to be disclosed below together with the accompanying drawings is intended to describe exemplary embodiments of the present invention, and is not intended to represent the only embodiments in which the present invention may be practiced. The following detailed description includes specific details to provide a thorough understanding of the present invention. However, those of ordinary skill in the art to which the present invention pertains knows that the present invention may be practiced without these specific details. In addition, throughout the specification, when a part is said to "comprising or including" a certain element, it does not exclude other elements, but may further include other elements unless otherwise stated. Means that. In addition, when it is determined that detailed descriptions of known functions or configurations may unnecessarily obscure the subject matter of the present invention in describing the embodiments of the present invention, detailed descriptions thereof will be omitted. In addition, terms to be described later are terms defined in consideration of functions in an embodiment of the present invention, which may vary according to the intention or custom of users or operators. Therefore, the definition should be made based on the contents throughout the present specification.

Configuration of deep learning learning image storage and processing device

1 is a block diagram showing a schematic configuration of a deep learning learning image storage and processing apparatus according to a first embodiment of the present invention, and FIG. 2 is a deep learning learning image storage and processing apparatus according to a second embodiment of the present invention. It is a block diagram showing the schematic configuration of.

Hereinafter, a schematic configuration of a deep learning learning image storage and processing apparatus according to the first and second embodiments of the present invention will be described with reference to FIGS. 1 to 2.

That is, it will be appreciated that the deep learning image storage and processing apparatus according to the first and second embodiments of the present invention may have additionally required configurations in addition to the schematic configurations shown in FIGS. 1 to 2.

<First Example>

In the deep learning learning image storage and processing apparatus according to the first embodiment of the present invention, referring to FIG. 1, a memory unit 10, an image region generation unit 20, a representative value extraction unit 30, and a learning image generation It is configured to include a section 40.

The memory unit 10 stores images for deep learning learning. The memory unit 10 stores images generated in various ways.

For a specific example, the image stored in the memory unit 10 is at least one of a microscope image collected through a microscope, a fluorescent dyed image dyed with a fluorescent substance through a fluorescent dyeing reagent, and a medical image captured from a medical device. I can.

The image region generation unit 20 generates at least one region from an image stored in the memory unit 10. For example, the image region generator 20 generates at least one image region using the number of bits and the number of pixels of the image. For a more specific example, the image region generator 20 generates at least one image region by dividing the number of bits of an image by a uniform number of pixels.

The representative value extraction unit 30 extracts a representative value from the image region by replacing the image region divided by the image region generating unit 20 with a specific coding. The representative value extracting unit 30 extracts a representative value using the number of bits of the image region replaced by a specific coding. For a more specific example, the representative value extracting unit 30 extracts the representative value through ^{2 n from the image region.} Here, n is the number of bits of the replaced image. That is, if the number of bits n of the replaced image is 2 bits, the representative value extracting unit 30 extracts four representative values from the image area.

In this way, in order for the representative value extraction unit 30 to extract the representative value, it is preferable that the image region generation unit 20 generate the image region equal to or greater than the representative value to be extracted.

Here, the representative value is a representative value of the image extracted from the image area through the representative value extraction unit 30, and the information value of the image to be generated as a learning image through a deep learning learning image storage and processing process to be described later. it means.

For example, the representative value extracting unit 30 extracts the representative value of the image area as a two-digit number or a four-digit number.

In addition, the representative value extracting unit 30 extracts a representative value so that representative values between adjacent image regions do not overlap with each other. For a specific example, when the representative value extracting unit 30 extracts the representative value from the first image area as a 2-digit number '01', the representative value is '01' in the second image area adjacent to the first image area. It is extracted with a 2-digit number other than'.

The learning image generation unit 40 generates a learning image for deep learning learning by arranging the representative values extracted by the representative value extraction unit 30. For example, the learning image generation unit 40 generates a learning image by arranging representative values in the order of image regions. For a specific example, the learning image generation unit 40 extracts the representative value from the representative value extraction unit 30 as '01', which is a two-digit number from the first image area, and '10', which is a two-digit number from the second image area. At this time, the representative values are arranged in the order of the image regions to generate a learning image with a value of '0110'.

<Second Example>

The deep learning learning image storage and processing apparatus according to the second embodiment of the present invention will be described with reference to FIG. 2, a memory unit 50, a pixel extraction unit 60, a representative value extraction unit 70, and a learning image generation unit. It is comprised of (80).

Here, since the memory unit 50 has the same configuration as the memory unit 10 of the first embodiment, a detailed description thereof will be omitted.

Unlike the conventional RGB system, the pixel extracting unit 60 extracts at least one pixel from an image stored in the memory unit 50. Specifically, the conventional RGB system extracts pixels from an image using colors of red, green, and blue, whereas the pixel extraction unit 60 extracts pixels from the image stored in the memory unit 10. Pixels are extracted for each fluorescence wavelength band. This pixel extracting unit 60 can extract more pixels from an image than in a conventional RGB system.

The representative value extracting unit 70 enlarges the pixel extracted by the pixel extracting unit 60 and then extracts the representative value. For example, when the representative value extracting unit 70 enlarges a pixel, the pixel may include information on a plurality of pixels (eg, 8 pieces). Thereafter, the representative value extracting unit 70 extracts, as a representative value, a pixel corresponding to the first two digits from the number of pixels less than half of the bits of the image stored according to the set conditional unit.

The learning image generation unit 80 arranges the representative values extracted by the representative value extraction unit 70 to generate a learning image for deep learning learning. For example, the learning image generator 80 may generate a learning image by arranging representative values in the order of the fluorescence wavelength bands of the images.

Deep learning learning image storage and processing method

3 is a step flow diagram of a deep learning learning image storage and processing method according to the first and second embodiments of the present invention, and FIG. 4 is a detailed step flow diagram of the representative value extraction step according to the first embodiment of the present invention. 5 is a view for explaining the process of generating an image area in the image area generation step according to the first embodiment of the present invention and a representative value extraction process in the representative value extraction step based on specific coding substitution, and FIG. 6 is a second diagram of the present invention. A detailed step flow diagram of a representative value extraction step according to an embodiment, and FIG. 7 is a diagram for explaining a process of extracting pixels of an image for each fluorescent wavelength band in the pixel extraction step according to the second embodiment of the present invention. A diagram for explaining the representative value extraction method in the conditional representative value extraction step and the representative value arrangement process in the representative value arrangement step according to the second embodiment of the present invention, and FIG. 9 is a representative value according to the second embodiment of the present invention. A diagram for explaining a representative value including an empty space generated through an arrangement step.

Hereinafter, a deep learning learning image storage and processing method according to the first embodiment of the present invention will be described with reference to FIGS. 3 to 5, and a second embodiment of the present invention with reference to FIGS. 3 and 6 to 9 We will explain how to store and process deep learning images according to an example.

<First Example>

First, the memory unit 10 stores an image 100 for deep learning learning (S10).

As described above, the image 100 stored in the memory unit 10 is a microscope image collected through a microscope, a fluorescent dyed image dyed with a fluorescent substance through a fluorescent dyeing reagent, and a medical image captured of a medical image of a medical device. It may be at least one of.

After the image storage step (S10), the image area generation unit 20 creates an image area in the image 100 stored in the memory unit 10, and the representative value extraction unit 30 obtains a representative value from the separated image area. Extract (S20).

In the representative value extraction step (S20), referring to FIG. 4, the image region generation unit 20 generates an image region (S21) and the representative value extraction unit 30 replaces the image region with a specific coding. It is subdivided into a step (S22) of extracting a representative value.

In the image region generation step (S21), referring to FIG. 5, the image region generation unit 20 divides the number of bits of the image 100 stored in the memory unit 10 into a uniform number of pixels, , 100c, 100d). Here, the image regions 100a, 100b, 100c, and 100d may be generated more or less than the illustrated ones.

In the representative value extraction step (S22), referring to FIG. 5, the representative value extraction unit 30 replaces the image regions 100a, 100b, 100c, and 100d with specific coding to extract a representative value from the image region 100. do. As described above, the representative value extracting unit 30 extracts ^{the representative value through 2 n} , where n is the number of bits of the image 100 replaced as described above. In addition, the representative value extracting unit 30 extracts the representative value in two or four digits, and the representative values between adjacent image regions 100a, 100b, 100c, and 100d do not overlap with each other.

As an example of this representative value extraction step (S22), when the number of bits (n) of the replaced image 100 is 2 bits, the representative value extraction unit 30 is 2 ^{n =} 4, Representative values may be extracted from the first, second, third, and fourth image regions 100a, 100b, 100c, and 100d. In addition, when the representative value extracting unit 30 extracts the representative value in two digits, the representative value extracting unit 30 is '00' from the first image area 100a and '01' from the second image area 100b. , The representative value may be extracted as '10' from the third image area 100c and '11' from the fourth image area 100d.

After the representative value extraction step (S22), the learning image generation unit 40 generates a learning image by arranging the representative values (S30). As described above, the learning image generator 40 may generate a learning image by arranging representative values in the order of the image regions 100a, 100b, 100c, and 100d, as described above.

Through the deep learning learning image storage and processing method according to the first embodiment of the present invention described so far, the deep learning learning image storage and processing apparatus according to the first embodiment of the present invention has a reduced capacity than the conventional image. By performing deep learning learning through images, the deep learning learning speed can be greatly improved.

<Second Example>

First, the memory unit 50 stores an image 200 for deep learning learning (S10).

On the other hand, the image 200 of the second exemplary embodiment differs only from the image 100 of the first exemplary embodiment, but refers to the same type of image.

After the image storage step (S10), the pixel extracting unit 60 extracts the pixels of the image 200 stored in the memory unit 50, and the representative value extracting unit 70 extracts a representative value from the pixels according to the set conditional part. Do (S20).

Referring to FIG. 6, the representative value extracting step S20 includes a step S23 of extracting a pixel from the image 200 stored in the memory unit 50 by the pixel extracting unit 60 and the representative value extracting unit 70. ) Is subdivided into a step (S24) of extracting a representative value from the pixel according to the set conditional part.

In the pixel extraction step S23, referring to FIG. 7, the pixel extraction unit 60 classifies the image 200 stored in the memory unit 50 by fluorescence wavelength band. For example, the image 200 is classified into a first image 200a, a second image 200b, a third image 200c, and a fourth image 200d by the pixel extracting unit 60. In this case, the first image 200a may be a red wavelength, the second image 200b may be a green wavelength, the third image 200c may be a blue wavelength, and the fourth image 200d may be a purple wavelength.

After that, the pixel extracting unit 60 includes a first pixel 201 from the classified first image 200a, a second pixel 202 from the second image 200b, and a third pixel from the third image 200c. (203), the fourth pixel 204 is extracted from the fourth image 200d, respectively. Here, the first, 2, 3, and 4 pixels 201, 202, 203, and 204 extracted by the pixel extraction unit 60 constitute the first, 2, 3, and 4 images 200a, 200b, 200c, and 200d. It means one pixel. When the first, second, third, and fourth pixels 201, 202, 203, and 204 are enlarged, a plurality (eg, eight) of pixel information may be formed.

After the pixel extraction step (S23), the conditional representative value extraction step (S24) will be described with reference to FIG. 8, from the first pixel 201 according to the set conditional part by the representative value extraction unit 70, the first representative value 201a ), a second representative value 202a from the second pixel 202, a third representative value 203a from the third pixel 203, and a fourth representative value 204a from the fourth pixel 204, respectively. .

Here, the conditional unit for extracting the representative value means a conditional unit for extracting a pixel corresponding to the first two digits from the number of pixels of less than half of the bits of the stored image 200 as a representative value.

After the conditional representative value extraction step (S24), the learning image generation unit 80 generates a learning image 300 by arranging the representative values extracted by the representative value extraction unit 70 (S30). Specifically, the learning image generation unit 80 arranges the first representative value 201a, the second representative value 202a, the third representative value 203a, and the fourth representative value 204a in the order of the fluorescent wavelength band to learn. An image 300 may be generated.

On the other hand, the learning image generation step (S30) looks at with reference to FIG. 9, when the learning image generation unit 80 generates the learning image 300 by arranging representative values, an empty space 400 in the learning image 300 ) May occur. In this case, the learning image generator 80 fills or removes the empty space 400 or adds image information to the empty space.

Here, the image information may be information including the number of fluorescent wavelengths of the image 200.

The deep learning learning image storage and processing apparatus according to the second embodiment of the present invention through the deep learning learning image storage and processing method according to the second embodiment of the present invention described so far is similar to the first embodiment of the present invention. , Deep learning learning speed can be greatly improved by performing deep learning learning through a learning image whose capacity is reduced compared to the existing image.

The above description of the present invention is for illustrative purposes only, and those of ordinary skill in the art to which the present invention pertains will be able to understand that other specific forms can be easily modified without changing the technical spirit or essential features of the present invention. will be. Therefore, it should be understood that the embodiments described above are illustrative and non-limiting in all respects. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present invention is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present invention. do.

10, 50: memory unit, 20: image area generation unit,
30, 70: representative value extraction unit, 40, 80: learning image generation unit,
60: pixel extraction unit, 100: image of the first embodiment,
100a: first image area, 100b: second image area,
100c: third image area, 100d: fourth image area,
200: the image of the second embodiment, 200a: the first image,
200b: second image, 200c: third image,
200d: fourth image, 201: first pixel,
201a: first representative value, 202: second pixel,
202a: second representative value, 203: third pixel,
203a: third representative value, 204: fourth pixel,
204a: fourth representative value, 300: arranged representative value,
400: empty space.

Claims (10)

A first step of capturing and storing at least one image by the memory unit;
A second step of extracting, by a pixel extracting unit, a pixel of the image stored in the memory unit, and by extracting a representative value from the pixel by a representative value extracting unit; And
A third step of generating a learning image by arranging the extracted representative values by a learning image generator; Including,
The second step,
The pixel extraction unit extracts pixels after classifying the images stored in the memory unit in order from images having a long fluorescent wavelength to images having a short fluorescent wavelength,
The representative value extracting unit extracts a pixel corresponding to the first two digits from the number of pixels of less than half of the bits of the stored image as a representative value,
The third step,
When the learning image generator generates the learning image, the generated empty space is filled or removed, or image information including the number of fluorescent wavelengths of the image is added to the empty space, and then the representative value and the image information are used. Deep learning learning image storage and processing method, characterized in that to generate the arranged learning image. The method of claim 1,
The first step,
Deep learning learning image storage and processing, characterized in that the image is at least one of a microscope image collected through a microscope, a fluorescent dyed image dyed with a fluorescent substance through a fluorescent dyeing reagent, and a medical image captured by a medical device Way. delete delete delete delete delete delete delete delete

Images