KR20240030176A - Gan-based artificial intelligence high-quality video compression system - Google Patents

Abstract

A GAN-based artificial function high-quality image compression system is disclosed. A GAN-based artificial function high-quality image compression system according to an embodiment of the present invention includes a data compression unit that compresses the input original image into data in a preset format; a data storage unit that stores data compressed in the data compression unit; a data restoration unit that retrieves the data stored in the data storage unit and generates a high-resolution restored image; and a data learning unit that trains at least one of the data compression unit and the data restoration unit.

Description

GAN-based artificial intelligence high-quality image compression system {GAN-BASED ARTIFICIAL INTELLIGENCE HIGH-QUALITY VIDEO COMPRESSION SYSTEM}

The present invention relates to an image compression system using a GAN network. More specifically, the original image is converted into a vector (Latent Vector) in the latent space and stored, and then the converted image is input into the GAN's generator network to convert the original image. It's about the return system.

An artificial intelligence system is a computer system that implements human-level intelligence, and unlike existing rule-based smart systems, it is a system in which machines learn, make decisions, and become smarter on their own. As artificial intelligence systems are used, the recognition rate improves and users' preferences can be more accurately understood, and existing rule-based smart systems are gradually being replaced by deep learning systems.

Artificial intelligence technology consists of machine learning (deep learning) and element technologies using machine learning. Machine learning is an algorithmic technology that classifies/learns the characteristics of input data on its own, and elemental technology is a technology that uses machine learning algorithms such as deep learning to mimic functions such as recognition and judgment of the human brain, including linguistic understanding and visual It consists of technical areas such as understanding, reasoning/prediction, knowledge expression, and motion control.

The various fields where artificial intelligence technology is applied are as follows. Linguistic understanding is a technology that recognizes and applies/processes human language/characters and includes natural language processing, machine translation, conversation systems, question and answer, and voice recognition/synthesis. Visual understanding is a technology that recognizes and processes objects like human vision, and includes object recognition, object tracking, image search, person recognition, scene understanding, spatial understanding, and image improvement. Inference/prediction is a technology that judges information to make logical inferences and predictions, and includes knowledge/probability-based reasoning, optimization prediction, preference-based planning, and recommendations. Knowledge expression is a technology that automatically processes human experience information into knowledge data, and includes knowledge construction (data creation/classification) and knowledge management (data use). Motion control is a technology that controls the autonomous driving of vehicles and the movement of robots, and includes motion control (navigation, collision, driving), operation control (behavior control), etc.

Meanwhile, as multimedia technology and computer technology develop, image generation technology using deep learning techniques is being developed, but there are problems in that the learning speed is significantly slow, computational complexity is high, and the resolution of the generated image is low.

In this regard, the prior art, Korean Patent Publication No. 10-2018-0004989, is a deep learning technology that combines a neural network circuit and a parallel processing processor to quickly perform the process of analyzing existing images and converting them into replacement images. Korea Patent Publication No. 10-2132690 discloses a high-resolution image restoration algorithm with the best restoration performance by applying face recognition pattern technology and a face recognition pattern technology to select ultra-high resolution image processing technology and method. We are disclosing a super-resolution image restoration system in which, when restoring an image and applying it to a monitoring system, the high-resolution restored image is similar to the original input image, and the high-resolution face recognition accuracy is closer to that of the original image.

However, the conventional high-resolution image restoration system has a disadvantage in that the computational processing process is complicated, and even when deep learning technology is applied, the amount of data used as learning data is large and requires a large amount of memory.

Therefore, research is needed on a GAN-based artificial intelligence high-quality image compression system that can compress and store image data and generate restored data at a level almost similar to the original data even if the compressed data is restored at the user's request.

The present invention applies a GAN-based artificial intelligence system to extract and compress feature information of image data, and restores the compressed image based on the extracted feature information, thereby enabling efficient processing, storage, and management of data. The purpose is to provide an artificial intelligence high-quality image compression system.

In addition, by applying the GAN algorithm consisting of a generator and a discriminator, two networks compete and learn to flexibly learn complex and diverse features of the original data, which is a GAN-based artificial intelligence that can generate restored data that approximates the original data. The purpose is to provide a high-quality image compression system.

The problems to be solved by the present invention are not limited to the problems mentioned above, and other problems to be solved by the present invention that are not mentioned herein can be explained to those skilled in the art from the description below. You will be able to understand it clearly.

The GAN-based artificial intelligence high-quality image compression system according to an embodiment of the present invention includes a data compression unit that compresses the input original image into data in a preset format, a data storage unit that stores the data compressed in the data compression unit, and a data storage unit. It includes a data restoration unit that calls the data stored in the unit and generates a high-resolution restored image, and a data learning unit that trains at least one of the data compression unit and the data restoration unit.

In addition, the data learning unit distinguishes the restored image by comparing the original image with the restored image generated by the generator and a generator that includes an encoder that compresses the original image into feature map data and a decoder that generates a high-resolution restored image from the feature map data. Includes a discriminator, wherein the encoder extracts a random variable corresponding to at least one feature information included in the original image and generates feature map data consisting of a combination of the random variables, and the decoder generates the generated through the encoder. A restored image is generated by applying a preset network layer to the feature map data, wherein the network layer includes a filter composed of a matrix of a preset size, and a feature map corresponding to at least one position in a multidimensional space preset by the filter. Response data of the data is calculated, and a restored image is generated based on the distribution of the response data.

In addition, the data learning unit learns a generator and a discriminator based on [Equation 1], where the discriminator learns so that the right side of [Equation 1] has the maximum value, and the generator learns the right side of [Equation 1] It is characterized by learning to have a minimum value.

[Equation 1]

(The D(x) is the transfer function of the discriminator, the G(x) is the transfer function of the generator, and the x~pdata(x) means data sampled from the probability distribution for actual data.)

In addition, the data learning unit is characterized in that learning is terminated when the discriminator cannot distinguish the restored image by comparing the restored image generated by the generator with the original image.

In addition, when the learning of the generator is completed, the data learning unit stores the encoder's original image compression process and the decoder's restored image generation process, applies the encoder's original image compression process to the data compression unit, and generates the decoder's restored image. The process is characterized by being applied to the data restoration department.

The GAN-based artificial intelligence high-quality image compression system of the present invention extracts and compresses feature information of image data by applying a GAN-based artificial intelligence system, and restores the compressed image based on the extracted feature information to process and store data. and management can be carried out efficiently.

In addition, by applying the GAN algorithm consisting of a generator and a discriminator, two networks can compete and learn to flexibly learn complex and diverse features of the original data, which has the effect of generating restored data that is close to the original data.

Figure 1 is a configuration diagram of a GAN-based artificial intelligence high-quality image compression system according to an embodiment of the present invention.
Figures 2 and 3 are diagrams for explaining the data learning unit of the GAN-based artificial intelligence high-quality image compression system according to an embodiment of the present invention.
Figure 4 is a diagram illustrating the GAN-based artificial intelligence high-quality image compression process according to an embodiment of the present invention.

Specific details, including the problem to be solved by the present invention, the means for solving the problem, and the effect of the invention, are included in the examples and drawings described below. The advantages and features of the present invention and methods for achieving them will become clear by referring to the embodiments described in detail below along with the accompanying drawings.

Hereinafter, the present invention will be described in more detail with reference to the attached drawings.

Figure 1 is a configuration diagram of a GAN-based artificial intelligence high-quality image compression system according to an embodiment of the present invention, and Figures 2 and 3 show the data learning unit of the GAN-based artificial intelligence high-quality image compression system according to an embodiment of the present invention. This is a diagram for explanation, and Figure 4 is a diagram for explaining the GAN-based artificial intelligence high-quality image compression process according to an embodiment of the present invention.

<Example 1>

Referring to Figure 1, the GAN-based artificial intelligence high-quality image compression system 100 according to an embodiment of the present invention includes a data compression unit 110, a data storage unit 120, a data restoration 130, and a data learning unit ( 140), wherein the data compression unit 110 compresses the input original image into data in a preset format, and the data storage unit 120 stores the data compressed in the data compression unit 110. , the data restoration 130 calls at least one of the compressed data stored in the data storage unit 120 and generates a high-resolution restored image for the compressed data, and the data learning unit 140 stores the data. At least one of the compression unit 110 and the data restoration unit 130 can be trained.

More specifically, as shown in FIGS. 2 and 3, the data learning unit 140 includes an encoder that compresses the original image into the feature map data and a decoder that generates a high-resolution restored image from the feature map data. It may include a generator 141 and a discriminator 142 that compares the restored image generated by the generator 141 with the original image to distinguish the restored image.

At this time, the encoder may extract a random variable corresponding to at least one feature information included in the original image and generate feature map data composed of a combination of the random variables.

For example, if the original image is a human image, the feature information may include eyebrow thickness, eye length, mouth size, etc., and if the original image is a pathology image, the size and texture of epidermal cells, nuclei, etc. , lesion area, etc. can be included.

Additionally, the original image has a statistical average value, and clusters of these average values can be distinguished as individual features. Therefore, the original image can be expressed as a combination of random variables corresponding to the individual features, and the encoder extracts feature information included in the original image and creates a feature composed of a combination of random variables corresponding to the feature information. Map data can be created.

At this time, the feature map data may be expressed as one point on a predetermined multidimensional feature space.

Meanwhile, the decoder may generate the restored image by applying a preset network layer to the feature map data generated through the encoder. At this time, the network layer may include a filter composed of a matrix of a preset size.

Accordingly, the decoder extracts key feature information by applying the filter to the image compressed in the form of feature map data, and generates a restored image based on the extracted key feature information.

When the filter is applied to the feature map data, response data of the feature map data corresponding to at least one location in a preset multidimensional space is calculated, and the restored image can be generated based on the distribution of the response data. .

Meanwhile, the data learning unit 140 calculates an error (loss) by comparing the original image and the restored image, and the error (loss) is calculated as the probability that the restored image is the original image, and the determination The generator 141 may update the weights of the network layer and the filter based on the error (loss) so that the generator 141 generates the restored image that the ruler 142 cannot distinguish from the original image.

In addition, as the weights of the network layer and the filter are updated, the generator 141 learns the process of generating the restored image, and the discriminator 142 distinguishes the restored image from the original image. However, based on the following [Equation 1], the image generator 130 learns the right side to have the minimum value, and the image determination unit 140 learns the right side to have the maximum value.

[Equation 1]

(The D(x) is the transfer function of the discriminator, the G(x) is the transfer function of the generator, and the x~pdata(x) means data sampled from the probability distribution for actual data.)

That is, the generator 141 may generate the restored image similar to the original image so that the discriminator 142 cannot accurately distinguish whether the restored image generated by the generator 141 is the original image.

More specifically, the first term on the right side of [Equation 1] can be omitted because it does not include G(x), which is the transfer function of the generator 141, and in order for the second term to have the minimum value, log(1-D Since (G(x)) must be minimal, D(G(x)) must be 1.

Accordingly, the generator 141 can learn to generate a restored image similar to the original image to the extent that the discriminator 142 can determine that the restored image generated by the generator 141 is the original image.

Meanwhile, the discriminator 142 can learn so that both the first and second terms on the right side of [Equation 1] have maximum values.

More specifically, the first term, log D(x), must be 1, and the second term, log(1-D(G(x))) must also be 1, so D(G(x)) must be 0.

In other words, the discriminator 142 must be able to more accurately distinguish between the original image and the restored image, and can learn to output 1 for the original image and 0 for the restored image.

Accordingly, the generator 141 learns to generate a restored image similar to the original image enough to deceive the discriminator 142, and the discriminator 142 uses the restored image generated by the generator 141. By learning to accurately distinguish, it is possible to generate a high-resolution restored image that is almost similar to the original image by the generator 141 and the discriminator 142.

Meanwhile, the data learning unit 140 may end learning when the discriminator 142 cannot distinguish the restored image by comparing the restored image generated by the generator 141 with the original image. there is.

In addition, when learning of the generator 141 is completed, the data learning unit 140 stores the original image compression process of the encoder and the restored image generation process of the decoder, and performs the original image compression process of the encoder. It is applied to the data compression unit 110, and the restored image generation process of the decoder can be applied to the data restoration unit 130.

That is, the data compression unit 110 and the data restoration unit 130 can compress and restore data more quickly by applying the algorithm learned in the data learning unit 140.

For example, when learning of the generator 141 is completed, the weights of the network layer and the filter applied by the generator 141 are determined, and when the weights of the network layer and the filter are determined, the generator 141 ) can generate the restored image that approximates the original image at once without going through any more learning process to update the weights of the network layer and the filter.

Referring to Figure 4, the GAN-based artificial intelligence high-quality image compression system 100 applies the GAN algorithm to learn the generator 141 and the discriminator 142 of the data learning unit 140, and the generator ( 141) and when the learning of the discriminator 142 is completed, the algorithm of the generator 141 is applied to the data compression unit 110 and the data restoration unit 130 to more quickly retrieve the original image 10. It is possible to compress and generate a high-resolution restored image 20.

<Example 2>

The GAN-based artificial intelligence high-quality image compression method according to an embodiment of the present invention includes a first step in which an original image is input, extracting a random variable corresponding to at least one feature information included in the original image, and extracting the feature information data. A second step of generating, a third step of converting the feature information data into feature map data displayed as a single point on a predetermined multidimensional feature space, a fourth step of generating a high-resolution restored image from the feature map data, and It may include a fifth step of comparing the restored image with the original image to determine the restored image.

In addition, the fourth step includes a step 4-1 of applying a predetermined network layer to the feature map data, a step 4-2 of calculating response data of the feature map data displayed on the multidimensional feature space, and the step 4-2 of applying a predetermined network layer to the feature map data. It may include a 4-3 step of generating the restored image based on the distribution of response data.

Therefore, by using the GAN-based artificial intelligence high-quality image compression method, a large-capacity, high-resolution original image is compressed into a feature map with a reduced area and a deepened depth by random variables and their combinations corresponding to the feature information included in the original image, It has the effect of making it easy to store and transmit data.

According to the effect of the present invention as described above, the feature information of image data is extracted and compressed by applying a GAN-based artificial intelligence system, and the compressed image is restored based on the extracted feature information to process, store, and manage the data. A GAN-based artificial intelligence high-quality image compression system and method that can perform efficiently can be provided.

In addition, by applying the GAN algorithm consisting of a generator and a discriminator, two networks compete and learn to flexibly learn complex and diverse features of the original data, which is a GAN-based artificial intelligence that can generate restored data that approximates the original data. A high-quality image compression system and method may be provided.

In addition, the purpose is to provide a GAN-based artificial intelligence high-quality image compression system that can learn about various evaluation criteria by continuously changing the discriminator that affects the error value that updates the generator's weight.

In addition, the generator provides a GAN-based artificial intelligence high-quality image compression system that can build a robust artificial intelligence engine that generates restored data that better approximates the original data by generating restored images that satisfy evaluation criteria from various perspectives. There is a purpose.

Additionally, the GAN-based artificial intelligence high-quality image compression method according to an embodiment of the present invention may be recorded on a computer-readable medium containing program instructions for performing various computer-implemented operations. The computer-readable medium may include program instructions, data files, data structures, etc., singly or in combination. The medium may have program instructions specifically designed and configured for the present invention, or may be known and available to those skilled in the art of computer software. Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes, optical media such as CD-ROMs and DVDs, and magnetic media such as floptical disks. -Includes optical media (magneto-optical media) and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, flash memory, etc. Examples of program instructions include machine language code, such as that produced by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc.

As described above, although one embodiment of the present invention has been described with limited examples and drawings, one embodiment of the present invention is not limited to the above-described embodiment, which is based on common knowledge in the field to which the present invention pertains. Anyone who has the knowledge can make various modifications and variations from this description. Accordingly, one embodiment of the present invention should be understood only by the scope of the claims set forth below, and all equivalent or equivalent modifications thereof shall fall within the scope of the spirit of the present invention.

110: data compression unit
120: data storage unit
130: Data restoration department
140: Data learning unit 141: Constructor
142: Discriminator

Claims (5)

a data compression unit that compresses the input original image into data in a preset format;
a data storage unit that stores data compressed in the data compression unit;
a data restoration unit that retrieves the data stored in the data storage unit and generates a high-resolution restored image; and
a data learning unit that trains at least one of the data compression unit and the data restoration unit;
A GAN-based artificial function high-quality image compression system including. According to paragraph 1,
The data learning department,
a generator including an encoder that compresses the original image into feature map data and a decoder that generates a high-resolution restored image from the feature map data; and
A discriminator that compares the restored image generated by the generator with the original image to distinguish the restored image,

The encoder is,
Extracting a random variable corresponding to at least one feature information included in the original image and generating feature map data composed of a combination of the random variables,

The decoder is,
The restored image is generated by applying a preset network layer to the feature map data generated through the encoder,
The network layer includes a filter composed of a matrix of a preset size,
Response data of the feature map data corresponding to at least one position in a preset multidimensional space is calculated by the filter,
A GAN-based artificial intelligence high-quality image compression system, characterized in that the restored image is generated based on the distribution of the response data. According to paragraph 2,
The data learning unit,
The generator and the discriminator are learned based on [Equation 1] below,

[Equation 1]

(The D(x) is the transfer function of the discriminator, the G(x) is the transfer function of the generator, and the x~pdata(x) means data sampled from the probability distribution for actual data.)

The discriminator is,
Learn so that the right side of [Equation 1] has the maximum value,
The constructor is:
A GAN-based artificial intelligence high-quality image compression system characterized by learning so that the right side of [Equation 1] has the minimum value. According to paragraph 2,
The data learning unit,
A GAN-based artificial intelligence high-quality image compression system, characterized in that learning is terminated when the discriminator cannot distinguish the restored image by comparing the restored image generated by the generator with the original image. According to clause 4,
The data learning department,
When learning of the generator is completed, the encoder's original image compression process and the decoder's restored image generation process are stored, the encoder's original image compression process is applied to the data compression unit, and the decoder's restored image is generated. A GAN-based artificial intelligence high-quality image compression system characterized by applying the process to the data restoration unit.