WO2023214633A1 - 초해상도 신경망을 기반으로 한 화질개선 방법 및 장치 - Google Patents
초해상도 신경망을 기반으로 한 화질개선 방법 및 장치 Download PDFInfo
- Publication number
- WO2023214633A1 WO2023214633A1 PCT/KR2022/019384 KR2022019384W WO2023214633A1 WO 2023214633 A1 WO2023214633 A1 WO 2023214633A1 KR 2022019384 W KR2022019384 W KR 2022019384W WO 2023214633 A1 WO2023214633 A1 WO 2023214633A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- distortion
- data set
- neural network
- learning
- super
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T5/00—Image enhancement or restoration
- G06T5/60—Image enhancement or restoration using machine learning, e.g. neural networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4053—Scaling of whole images or parts thereof, e.g. expanding or contracting based on super-resolution, i.e. the output image resolution being higher than the sensor resolution
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T3/00—Geometric image transformations in the plane of the image
- G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting
- G06T3/4046—Scaling of whole images or parts thereof, e.g. expanding or contracting using neural networks
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING OR CALCULATING; COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
Definitions
- This disclosure relates to a method and device for improving image quality based on a super-resolution neural network.
- low-quality images can be converted into high-quality images.
- a picture quality improvement model can be learned based on machine learning based on artificial neural networks.
- sufficient supervised learning data for distortion is required.
- distortion refers to blur and noise found in low-quality images.
- As a method of collecting supervised learning data there is a method of manually converting low-quality data into high-quality data. This method consumes considerable cost. Meanwhile, depending on the service environment that provides video, there are cases where the supervised learning data itself cannot be collected.
- the prior art collected supervised learning data by randomly adding distortions such as blur, noise, and compression to the target image.
- This method is a universal method that can be applied to all images, but it has the problem of generating supervised learning data that does not focus on the distortion characteristics reflected only in the target image.
- the image quality improvement model learned based on a data set reflecting many distortion characteristics has a problem in that it is not optimized for converting the target image into a high-definition image because it does not sufficiently learn the distortion characteristics reflected only in the target image.
- the image quality improvement device calculates the similarity between the distortion characteristic value of the learning data set and the distortion characteristic value of the service data set, thereby generating the distortion most similar to the distortion characteristic of the service data set of the target image. You can select the added learning data set.
- an image quality improvement apparatus may select a super-resolution neural network optimized for specific distortion characteristics based on similarity between distortion characteristic values.
- an image quality improvement method optimized for the distortion characteristics of a target image includes the process of generating one or more training data sets to which one or more distortions are added; Inputting the one or more learning data sets into a Degradation Encoder Neural Network (DEN) to obtain each learning distortion characteristic value; Inputting a service data set consisting of image patches of the target image into the distortion coding neural network to obtain service distortion characteristic values; A process of calculating similarity between each learning distortion characteristic value and the service distortion characteristic value; and a process of selecting a learning data set with the highest similarity to the service distortion characteristic value.
- DEN Degradation Encoder Neural Network
- one or more super resolution neural networks are each used using learning data sets with different distortions added.
- a process of learning to optimize for specific distortion A process of calculating similarity between a learning data set and a service data set applied to each of the one or more super-resolution neural networks using a distortion encoding neural network (DEN);
- DEN distortion encoding neural network
- a process of selecting a super-resolution neural network learned from a learning data set with the highest similarity among the one or more super-resolution neural networks and a method of improving image quality including a process of converting a target image into a high-definition image using a selected super-resolution neural network.
- an image quality improvement device optimized for distortion characteristics of a target image includes: a memory storing one or more commands; And a processor, wherein the processor optimizes each of one or more Super Resolution Neural Networks (SRNs) to a specific distortion by executing the one or more instructions, using training data sets with different distortions added thereto. Learn as much as possible, and calculate the similarity between the training data set and the service data set applied to each of the one or more super-resolution neural networks using a distortion encoding neural network (DEN: Degradation Encoder Neural Network), and calculate the similarity among the one or more super-resolution neural networks.
- a super-resolution neural network learned with the highest learning data set is selected, and an image quality improvement device is provided that converts the target image into a high-definition image using the selected super-resolution neural network.
- the target image can be converted into a high-definition image using a super-resolution neural network optimized for the target image.
- a super-resolution neural network is learned by focusing on the distortion characteristics of the target image among the previously learned super-resolution neural networks by calculating the similarity between the distortion characteristics of the learning data set and the distortion characteristics of the service data set.
- Figure 1A is a flowchart of a method for improving image quality according to an embodiment of the present disclosure.
- Figure 1B is a flowchart of a method for improving image quality according to another embodiment of the present disclosure.
- Figure 2 is an example diagram illustrating a process for generating a learning data set according to an embodiment of the present disclosure.
- Figure 3 is an exemplary diagram illustrating a process of learning a super-resolution neural network according to an embodiment of the present disclosure.
- FIG. 4 is an exemplary diagram illustrating a process of extracting distortion characteristics by inputting a service data set into a distortion coding neural network according to an embodiment of the present disclosure.
- FIG. 5A is an example diagram illustrating a process of calculating a weight based on the similarity between a learning data set and a service data set according to an embodiment of the present disclosure.
- FIG. 5B is an exemplary diagram illustrating an example of a learning data set according to an embodiment of the present disclosure.
- Figure 6 is a block diagram of an image quality improvement device according to an embodiment of the present disclosure.
- symbols such as first, second, i), ii), a), and b) may be used. These codes are only used to distinguish the component from other components, and the nature, sequence, or order of the component is not limited by the code. In the specification, when a part is said to 'include' or 'have' a certain element, this means that it does not exclude other elements, but may further include other elements, unless explicitly stated to the contrary. .
- Figure 1A is a flowchart of a method for improving image quality according to an embodiment of the present disclosure.
- the image quality improvement device can generate one or more training data sets to which one or more distortions have been added (S100).
- distortion includes blur, noise, etc.
- the method of adding distortion may be to add one distortion characteristic at a specific ratio, or to add two or more distortion characteristics in combination at a certain ratio.
- a training data set is one or more image patches included in a training video classified into one training set.
- the image quality improvement device may calculate the similarity (score) between the distortion characteristic value of the learning data set and the distortion characteristic value of the service data set (S102).
- the service data set is one or more image patches constituting the target image subject to image quality improvement classified into one service set.
- An image quality improvement device can acquire distortion characteristic values of each data set using a distortion encoding neural network (DEN: Degradation Encoder Neural Network).
- the distortion coding neural network is a neural network that outputs the characteristic value of the distortion included in the input image.
- the output value of the distortion coding neural network is a vector that is clustered according to the distortion characteristics and intensity of the input image.
- Distortion coding neural networks can be learned based on contrastive learning.
- Contrast learning is a method of learning a learning object using an objective function designed to minimize the loss between image patches with the same distortion and maximize the loss between image patches with different distortions.
- Contrast learning uses the first objective function ( ) is used to learn the distortion coding neural network.
- the first objective function can be expressed using Equation 1.
- DEN( ⁇ ) represents the output of the distortion coding neural network, is the total number of different distortions to be learned, and are a query sample and a positive sample corresponding to the mth distortion-added image patch, respectively.
- An image patch refers to one image among a plurality of images constituting one data set. is a negative sample corresponding to the non-mth distortion-added image patch, is the total number of negative samples.
- the image quality improvement device may input one or more learning data sets and one or more service data sets into a distortion coding neural network.
- the image quality improvement device may obtain the distortion characteristic value of each of one or more learning data sets and the distortion characteristic value of the service data set based on the output of the distortion coding neural network.
- Equation 2 The similarity between the distortion characteristic value of an arbitrary learning data set and the distortion characteristic value of the service data set can be calculated using Equation 2.
- samples representing each data set can be extracted in advance.
- Methods for extracting samples representing a data set include randomly sampling from the data set or selecting representative images for each cluster using K-means clustering.
- the average of the distortion characteristic values for at least one sample selected from each learning data set is Can be used as a value expressing the distortion characteristic of each learning data set (hereinafter referred to as the learning distortion characteristic value).
- These learning distortion characteristic values can be interpreted as values expressing the distortion characteristics that can best be restored by a super-resolution neural network to be learned using the corresponding learning data set.
- the average of the distortion characteristic values for at least one sample selected from the service set is Can be used as a value expressing the distortion characteristics of the service data set (hereinafter, service distortion characteristic value).
- the image quality improvement device may pre-calculate learning distortion characteristic values for each of one or more learning data sets and generate a look-up table in which the calculated values are arranged in the form of a table.
- the video quality improvement device can select a learning data set with the highest similarity to the service distortion characteristic value (S104).
- the reason for selecting the learning data set with the highest similarity is that a super-resolution neural network can be trained to optimize the distortion characteristics of the target image using this learning data set.
- the image quality improvement device can learn a super resolution neural network (SRN) based on the selected learning data set (S106).
- SRN super resolution neural network
- the image quality improvement device can convert the target image into a high-definition image using a super-resolution neural network learned based on the selected learning data set (S108).
- Figure 1B is a flowchart of a method for improving image quality according to another embodiment of the present disclosure.
- the image quality improvement device may have one or more super-resolution neural networks optimized for different distortions (S150).
- An image quality improvement device can generate one or more learning data sets optimized for different distortions.
- the image quality improvement device can train one or more super-resolution neural networks using this learning data set.
- Each of the one or more super-resolution neural networks corresponds to a super-resolution neural network optimized for a specific distortion.
- the image quality improvement device may use a distortion coding neural network to calculate the similarity between each learning data set applied to one or more super-resolution neural networks and the service data set (S152).
- a distortion coding neural network uses a distortion coding neural network to provide a learning distortion characteristic value, which is a value expressing the distortion characteristics of the learning data set applied to each super-resolution neural network, and a value expressing the distortion characteristics of the service data set.
- Service distortion characteristic values can be obtained.
- the learning distortion characteristic value is the average of the output of the distortion coding neural network for at least one sample selected from each learning data set
- the service distortion characteristic value is the distortion for at least one sample selected from the service data set. It may be the average of the output of the encoding neural network.
- the image quality improvement device may pre-calculate the learning distortion characteristic value of the learning data set applied to each super-resolution neural network and store it in the form of a lookup table.
- the image quality improvement device may calculate similarity based on the difference between the learning distortion characteristic value and the service distortion characteristic value corresponding to each super-resolution neural network.
- the image quality improvement device may select a super-resolution neural network to which a learning data set with the highest similarity to the service data set is applied among one or more pre-trained super-resolution neural networks (S154).
- the image quality improvement device can convert the target image into a high-definition image using the selected super-resolution neural network (S156).
- Figure 2 is an example diagram illustrating a process for generating a learning data set according to an embodiment of the present disclosure.
- the image quality improvement device may add distortion to the original training image 200.
- the image quality improvement device includes a learning data set (202) in which blur is added to the original learning image (200), a learning data set (204) in which noise is added to the original image (200), and a learning data set in which both blur and noise are added. (206) can be generated.
- the image quality improvement device can generate a plurality of learning data sets 202, 204, and 206 by combining blur and noise at various intensities.
- An image quality improvement device can learn a super-resolution neural network based on a learning data set that focuses on the distortion characteristics of the target image by generating a learning data set with distortion similar to the distortion characteristics of the target image.
- Figure 3 is an exemplary diagram illustrating a process of learning a super-resolution neural network according to an embodiment of the present disclosure.
- the image quality improvement device inputs a training image 300 to which random distortion has been added into the distortion coding neural network 30, obtains a learning distortion characteristic value 302, and
- the output image 306 can be obtained by inputting it into the resolution neural network 32.
- the image quality improvement device may calculate a second objective function based on the difference between the output image 306 and the target image 308.
- the second objective function is a function for training a super-resolution neural network (SRN).
- SRN super-resolution neural network
- the weight 304 calculated based on the learning distortion characteristic value 302 may be reflected in the second objective function.
- the second objective function can be expressed as Equation 3.
- SRN ( ) is the output image 306 of the distortion coding neural network
- the high-definition target image 308 that is the target of image quality improvement
- the subscript i means the ith sample in the learning batch. means weight.
- the weight may be determined based on the similarity between the training image 300 of the training data set and the target image of the service data set, and can be expressed using Equation 4.
- Methods for extracting samples include randomly sampling from the service data set or selecting representative images for each cluster using the K-means clustering algorithm.
- Video quality improvement devices are used for specific service data sets. may be calculated in advance and stored in a lookup table.
- the image quality improvement device can learn a super-resolution neural network using a weight 304 and a second objective function generated based on calculation of the difference between the output image 306 and the target image 308.
- the similarity between the learning data set and the service data set may be reflected as a weight in the error backpropagation process.
- FIG. 4 is an exemplary diagram illustrating a process of extracting distortion characteristics by inputting a service data set into a distortion coding neural network according to an embodiment of the present disclosure.
- the image quality improvement device inputs samples 400 extracted from the service data set into a distortion coding neural network before training the super-resolution neural network to obtain the service distortion characteristic value 402 in advance.
- the service distortion characteristic value (402) is It can be expressed as a vector value (N is a natural number).
- the video quality improvement device may calculate the average 404 of the service distortion characteristic values based on each service distortion characteristic value 402.
- the average 404 of the service distortion characteristic value may be a value representing the distortion characteristic of the service data set.
- FIG. 5A is an example diagram illustrating a process of calculating a weight based on the similarity between a learning data set and a service data set according to an embodiment of the present disclosure.
- FIG. 5B is an exemplary diagram illustrating an example of a learning data set according to an embodiment of the present disclosure.
- the image quality improvement device may input each image patch of the learning data set 500 into a distortion coding neural network to obtain a learning distortion characteristic value 502 of each image.
- different distortions may be added to the image patches of the training data set 500, as shown in FIG. 5B.
- the distortion characteristic value for learning (502) is It can be expressed as a vector value (N is a natural number).
- the image quality improvement device may calculate the weight 506 of the learning data set based on the similarity between the learning distortion characteristic value 502 and the average service distortion characteristic value 504.
- similarity can be utilized in two ways.
- the first method as described in Figure 3, calculates a weight for each sample in the learning data set based on similarity and reflects it in the error backpropagation process for the super-resolution neural network.
- the second way to utilize similarity is to select one of a plurality of super-resolution neural networks or a plurality of learning data sets based on similarity, as described above with reference to FIGS. 1A and 1B.
- the data to be provided to the service is specified, there is no need to train multiple super-resolution neural networks, and it may be efficient to train only one super-resolution neural network based on the training data set with the highest similarity, as shown in Figure 1a.
- the similarity of each of the plurality of data sets may be calculated. By comparing the calculated similarities, the data set with the highest similarity can be selected to train a super-resolution neural network.
- Figure 6 is a block diagram of an image quality improvement device according to an embodiment of the present disclosure.
- the image quality improvement device 600 includes all or part of a memory 602 and a processor 604.
- the memory 602 may store a program that performs a method for improving image quality according to an embodiment of the present invention.
- a program may include a plurality of instructions executable by the processor 604, and the image quality improvement method can be performed by executing the plurality of instructions by the processor 604.
- the memory 602 may include at least one of volatile memory and non-volatile memory.
- Volatile memory includes Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM), and non-volatile memory includes flash memory.
- SRAM Static Random Access Memory
- DRAM Dynamic Random Access Memory
- Processor 604 may include at least one component capable of executing at least one instruction.
- the processor 604 can execute instructions stored in the memory 602, and can perform the image quality improvement method according to the present disclosure by executing the instructions.
- Various implementations of the devices and methods described herein may be implemented as digital electronic circuits, integrated circuits, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), computer hardware, firmware, software, and/or combinations thereof. It can be realized.
- FPGAs field programmable gate arrays
- ASICs application specific integrated circuits
- Various implementations of the systems and techniques described herein may include digital electronic circuits, integrated circuits, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), computer hardware, firmware, software, and/or these. It can be realized through combination.
- These various implementations may include being implemented as one or more computer programs executable on a programmable system.
- the programmable system includes at least one programmable processor (which may be a special purpose processor) coupled to receive data and instructions from and transmit data and instructions to a storage system, at least one input device, and at least one output device. or may be a general-purpose processor).
- Computer programs also known as programs, software, software applications or code
- Computer-readable recording media include all types of recording devices that store data that can be read by a computer system. These computer-readable recording media are non-volatile or non-transitory such as ROM, CD-ROM, magnetic tape, floppy disk, memory card, hard disk, magneto-optical disk, and storage device. It may be a medium, and may further include a transitory medium such as a data transmission medium. Additionally, the computer-readable recording medium may be distributed in a computer system connected to a network, and the computer-readable code may be stored and executed in a distributed manner.
- a programmable computer includes a programmable processor, a data storage system (including volatile memory, non-volatile memory, or another type of storage system, or a combination thereof), and at least one communication interface.
- a programmable computer may be one of a server, network device, set-top box, embedded device, computer expansion module, personal computer, laptop, personal data assistant (PDA), cloud computing system, or mobile device.
- PDA personal data assistant
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Image Processing (AREA)
- Compression Or Coding Systems Of Tv Signals (AREA)
Abstract
Description
Claims (10)
- 타겟 영상의 왜곡 특성에 최적화된 화질개선 방법으로서,하나 이상의 왜곡을 첨가한 학습용 데이터 세트를 하나 이상 생성하는 과정;상기 하나 이상의 학습용 데이터 세트를 왜곡 부호화 신경망(DEN: Degradation Encoder Neural Network)에 입력하여 각각의 학습용 왜곡 특성값을 획득하는 과정;상기 타겟 영상의 이미지 패치로 구성된 서비스 데이터 세트를 상기 왜곡 부호화 신경망에 입력하여 서비스 왜곡 특성값을 획득하는 과정;상기 각각의 학습용 왜곡 특성값과 상기 서비스 왜곡 특성값 간의 유사도를 계산하는 과정; 및상기 서비스 왜곡 특성값과 가장 유사도가 높은 학습용 데이터 세트를 선정하는 과정을 포함하는 영상 화질개선 방법.
- 제1항에 있어서,상기 선정된 학습용 데이터 세트를 기초로 하나의 초해상도 신경망(SRN: Super Resolution Neural Network)을 학습시키는 과정; 및상기 초해상도 신경망을 이용하여, 상기 타겟 영상을 고화질 영상으로 변환하는 과정을 더 포함하는 영상 화질개선 방법.
- 제1항에 있어서,상기 학습용 왜곡 특성값은, 각 학습용 데이터 세트에서 선정된 적어도 하나의 샘플들에 대한 상기 왜곡 부호화 신경망의 출력의 평균이고,상기 서비스 왜곡 특성값은, 상기 서비스 데이터 세트에서 선정된 적어도 하나의 샘플들에 대한 상기 왜곡 부호화 신경망의 출력의 평균인 것을 특징으로 하는, 영상 화질개선 방법.
- 제1항에 있어서,상기 유사도는, 상기 각각의 학습용 왜곡 특성값과 상기 서비스 왜곡 특성값 간의 차를 기초로 산출되는 것을 특징으로 하는, 영상 화질개선방법.
- 명령어가 저장된 컴퓨터로 읽을 수 있는 기록매체로서, 상기 명령어는 상기 컴퓨터에 의해 실행될 때 상기 컴퓨터로 하여금, 제1항에 따른 방법을 수행하도록 하는, 컴퓨터로 읽을 수 있는 기록매체.
- 타겟 영상의 왜곡 특성에 최적화된 화질개선 방법으로서,서로 다른 왜곡이 첨가된 학습용 데이터 세트를 이용하여, 하나 이상의 초해상도 신경망(SRN: Super Resolution Neural Network) 각각을 특정한 왜곡에 최적화되도록 학습시키는 과정;왜곡 부호화 신경망(DEN: Degradation Encoder Neural Network)을 이용하여, 상기 하나 이상의 초해상도 신경망 각각에 적용된 학습용 데이터 세트와 서비스 데이터 세트의 유사도를 계산하는 과정;상기 하나 이상의 초해상도 신경망 중 유사도가 가장 높은 학습용 데이터 세트로 학습된 초해상도 신경망을 선정하는 과정; 및선정된 초해상도 신경망을 이용하여 타겟 영상을 고화질 영상으로 변환하는 과정을 포함하는 영상 화질개선 방법.
- 제6항에 있어서,상기 유사도를 계산하는 과정은,상기 왜곡 부호화 신경망을 이용하여, 각 초해상도 신경망에 적용된 학습용 데이터 세트의 왜곡의 특성을 표현하는 값인 학습용 왜곡 특성값 및 상기 서비스 데이터 세트의 왜곡의 특성을 표현하는 값인 서비스 왜곡 특성값을 획득하는 과정; 및상기 학습용 왜곡 특성값과 상기 서비스 왜곡 특성값 간의 차를 기초로, 상기 유사도를 계산하는 과정을 포함하는 것을 특징으로 하는, 영상화질 개선방법.
- 제7항에 있어서,상기 학습용 왜곡 특성값은, 각 학습용 데이터 세트에서 선정된 적어도 하나의 샘플들에 대한 상기 왜곡 부호화 신경망의 출력의 평균이고,상기 서비스 왜곡 특성값은, 상기 서비스 데이터 세트에서 선정된 적어도 하나의 샘플들에 대한 상기 왜곡 부호화 신경망의 출력의 평균인 것을 특징으로 하는, 영상 화질개선 방법.
- 명령어가 저장된 컴퓨터로 읽을 수 있는 기록매체로서, 상기 명령어는 상기 컴퓨터에 의해 실행될 때 상기 컴퓨터로 하여금, 제6항에 따른 방법을 수행하도록 하는, 컴퓨터로 읽을 수 있는 기록매체.
- 타겟 영상의 왜곡 특성에 최적화된 화질개선 장치로서,하나 이상의 명령어를 저장하는 메모리; 및프로세서를 포함하되,상기 프로세서는, 상기 하나 이상의 명령어를 실행함으로써,서로 다른 왜곡이 첨가된 학습용 데이터 세트를 이용하여, 하나 이상의 초해상도 신경망(SRN: Super Resolution Neural Network) 각각을 특정한 왜곡에 최적화되도록 학습시키고,왜곡 부호화 신경망(DEN: Degradation Encoder Neural Network)을 이용하여, 상기 하나 이상의 초해상도 신경망 각각에 적용된 학습용 데이터 세트와 서비스 데이터 세트의 유사도를 계산하고,상기 하나 이상의 초해상도 신경망 중 유사도가 가장 높은 학습용 데이터 세트로 학습된 초해상도 신경망을 선정하며,선정된 초해상도 신경망을 이용하여 타겟 영상을 고화질 영상으로 변환하는,영상 화질개선 장치.
Priority Applications (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202280097559.3A CN119452391A (zh) | 2022-05-06 | 2022-12-01 | 基于超分辨率神经网络提高图像质量的方法和装置 |
| US18/938,248 US20250061547A1 (en) | 2022-05-06 | 2024-11-05 | Method and device for improving image quality on basis of super-resolution neural network |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| KR10-2022-0056193 | 2022-05-06 | ||
| KR1020220056193A KR20230156585A (ko) | 2022-05-06 | 2022-05-06 | 초해상도 신경망을 기반으로 한 화질개선 방법 및 장치 |
Related Child Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| US18/938,248 Continuation US20250061547A1 (en) | 2022-05-06 | 2024-11-05 | Method and device for improving image quality on basis of super-resolution neural network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2023214633A1 true WO2023214633A1 (ko) | 2023-11-09 |
Family
ID=88646560
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/KR2022/019384 Ceased WO2023214633A1 (ko) | 2022-05-06 | 2022-12-01 | 초해상도 신경망을 기반으로 한 화질개선 방법 및 장치 |
Country Status (4)
| Country | Link |
|---|---|
| US (1) | US20250061547A1 (ko) |
| KR (1) | KR20230156585A (ko) |
| CN (1) | CN119452391A (ko) |
| WO (1) | WO2023214633A1 (ko) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| EP4328848B1 (en) * | 2022-08-22 | 2026-05-20 | Sony Interactive Entertainment Europe Ltd | INCREASING THE SCALE OF IMAGE DATA |
| US12602748B2 (en) * | 2023-05-23 | 2026-04-14 | Constructor Technology Ag | Automatically enhancing image quality in machine learning training dataset by using deep generative models |
| WO2025230108A1 (ko) * | 2024-04-30 | 2025-11-06 | 삼성전자 주식회사 | 입력 이미지보다 화질이 향상된 고화질 이미지를 생성하는 방법 및 이를 위한 전자 장치 |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101922964B1 (ko) * | 2017-06-27 | 2018-11-28 | 아주대학교산학협력단 | 이미지 왜곡 검출을 이용한 이미지 복원 장치 및 방법 |
| KR20200127766A (ko) * | 2019-05-03 | 2020-11-11 | 삼성전자주식회사 | 영상 처리 장치 및 그 영상 처리 방법 |
| JP2021502644A (ja) * | 2017-11-09 | 2021-01-28 | 京東方科技集團股▲ふん▼有限公司Boe Technology Group Co.,Ltd. | 画像処理方法、処理装置及び処理デバイス |
| KR102388588B1 (ko) * | 2020-06-12 | 2022-04-20 | 베이징 시아오미 파인콘 일렉트로닉스 컴퍼니 리미티드 | 이미지 노이즈 제거 모델의 트레이닝 방법, 이미지 노이즈 제거 방법, 장치 및 매체 |
| JP2022064389A (ja) * | 2020-10-14 | 2022-04-26 | プラスマン合同会社 | 画像処理装置、画像処理方法およびプログラム |
-
2022
- 2022-05-06 KR KR1020220056193A patent/KR20230156585A/ko active Pending
- 2022-12-01 CN CN202280097559.3A patent/CN119452391A/zh active Pending
- 2022-12-01 WO PCT/KR2022/019384 patent/WO2023214633A1/ko not_active Ceased
-
2024
- 2024-11-05 US US18/938,248 patent/US20250061547A1/en active Pending
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| KR101922964B1 (ko) * | 2017-06-27 | 2018-11-28 | 아주대학교산학협력단 | 이미지 왜곡 검출을 이용한 이미지 복원 장치 및 방법 |
| JP2021502644A (ja) * | 2017-11-09 | 2021-01-28 | 京東方科技集團股▲ふん▼有限公司Boe Technology Group Co.,Ltd. | 画像処理方法、処理装置及び処理デバイス |
| KR20200127766A (ko) * | 2019-05-03 | 2020-11-11 | 삼성전자주식회사 | 영상 처리 장치 및 그 영상 처리 방법 |
| KR102388588B1 (ko) * | 2020-06-12 | 2022-04-20 | 베이징 시아오미 파인콘 일렉트로닉스 컴퍼니 리미티드 | 이미지 노이즈 제거 모델의 트레이닝 방법, 이미지 노이즈 제거 방법, 장치 및 매체 |
| JP2022064389A (ja) * | 2020-10-14 | 2022-04-26 | プラスマン合同会社 | 画像処理装置、画像処理方法およびプログラム |
Also Published As
| Publication number | Publication date |
|---|---|
| KR20230156585A (ko) | 2023-11-14 |
| US20250061547A1 (en) | 2025-02-20 |
| CN119452391A (zh) | 2025-02-14 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2021107610A1 (en) | Method and system for generating a tri-map for image matting | |
| WO2018217019A1 (ko) | 신경망 학습 기반의 변종 악성 코드를 탐지하기 위한 장치, 이를 위한 방법 및 이 방법을 수행하기 위한 프로그램이 기록된 컴퓨터 판독 가능한 기록매체 | |
| CN113052868B (zh) | 一种抠图模型训练、图像抠图的方法及装置 | |
| WO2021225294A1 (ko) | 샘플 기반 정규화 기법을 이용한 전이 학습장치 및 방법 | |
| CN111898544B (zh) | 文字图像匹配方法、装置和设备及计算机存储介质 | |
| WO2021098618A1 (zh) | 数据分类方法、装置、终端设备及可读存储介质 | |
| WO2020045714A1 (ko) | 콘텐츠 인식 방법 및 시스템 | |
| KR20230156585A (ko) | 초해상도 신경망을 기반으로 한 화질개선 방법 및 장치 | |
| WO2017138766A1 (ko) | 하이브리드 기반의 영상 클러스터링 방법 및 이를 운용하는 서버 | |
| CN114863221A (zh) | 检测模型的训练方法、装置、系统、设备及存储介质 | |
| WO2021172674A1 (ko) | 재귀 그래프 모델링을 통한 비디오 요약 생성 장치 및 방법 | |
| WO2022260392A1 (ko) | 단말에서 동작하는 이미지 프로세싱 인공 신경망 모델 생성 방법 및 시스템 | |
| CN111353536A (zh) | 图像的标注方法、装置、可读介质和电子设备 | |
| WO2023249372A1 (ko) | 딥러닝 모델의 학습을 위한 의료 데이터의 처리 방법, 프로그램 및 장치 | |
| CN114896594B (zh) | 基于图像特征多注意力学习的恶意代码检测装置及方法 | |
| CN114969412A (zh) | 一种人脸图片归档方法、装置以及计算机可读存储介质 | |
| WO2022092445A1 (ko) | 전자 장치 및 그 제어 방법 | |
| WO2019147020A1 (ko) | 이미지의 품질 향상을 위하여 이미지를 처리하는 방법 및 장치 | |
| CN118230066A (zh) | 图像分类方法、图像认证方法、装置、介质、设备及产品 | |
| WO2023249184A1 (ko) | 적대적 학습 시스템 및 적대적 학습 방법 | |
| WO2025041871A1 (ko) | 딥러닝 네트워크 아키텍처 인코딩과 학습 중 양자화 파라미터 예측 모델 기반 양자화 방법 및 장치 | |
| WO2020175729A1 (ko) | 가우시안 특징점맵과 회귀 기법을 이용한 얼굴 특징점 검출 장치 및 방법 | |
| CN117315310A (zh) | 一种图像识别方法、图像识别模型训练方法及装置 | |
| WO2011007970A1 (en) | Method and apparatus for processing image | |
| WO2023113437A1 (ko) | 메모리를 이용하는 의미론적 영상 분할 장치 및 방법 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 22940872 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 202280097559.3 Country of ref document: CN |
|
| WWP | Wipo information: published in national office |
Ref document number: 202280097559.3 Country of ref document: CN |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 22940872 Country of ref document: EP Kind code of ref document: A1 |


