WO2020006762A1 - 训练图像修复模型的方法、图像修复方法、装置、介质及设备 - Google Patents

训练图像修复模型的方法、图像修复方法、装置、介质及设备 Download PDF

Info

Publication number
WO2020006762A1
WO2020006762A1 PCT/CN2018/094900 CN2018094900W WO2020006762A1 WO 2020006762 A1 WO2020006762 A1 WO 2020006762A1 CN 2018094900 W CN2018094900 W CN 2018094900W WO 2020006762 A1 WO2020006762 A1 WO 2020006762A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
loss value
repair
repair model
missing
Prior art date
Application number
PCT/CN2018/094900
Other languages
English (en)
French (fr)
Inventor
胡欢
刘兆祥
廉士国
Original Assignee
深圳前海达闼云端智能科技有限公司
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by 深圳前海达闼云端智能科技有限公司 filed Critical 深圳前海达闼云端智能科技有限公司
Priority to CN201880001170.8A priority Critical patent/CN108885787B/zh
Priority to PCT/CN2018/094900 priority patent/WO2020006762A1/zh
Publication of WO2020006762A1 publication Critical patent/WO2020006762A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T5/00—Image enhancement or restoration
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR 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; CALCULATING OR 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

Definitions

  • the present disclosure relates to the field of image processing, and in particular, to a method for training an image repair model, an image repair method, a device, a medium, and a device.
  • image restoration is to restore the missing part of the image, which has been widely used in real life.
  • image restoration is usually implemented by deep learning methods, such as obtaining an image restoration model by training a self-encoding full convolutional neural network or an adversarial generation network, thereby obtaining a restored image based on the image restoration model.
  • the training is supervised only by calculating the pixel difference between the network repaired image and the complete image, so that the trained image repair model has low robustness.
  • An object of the present disclosure is to provide a method, an image repair method, a device, a medium, and a device for training an image repair model that are widely used.
  • a method for training an image repair model includes:
  • the image repair model is updated according to the loss value.
  • an image repair method including:
  • the image to be repaired is input to an image repair model to obtain a repaired image, wherein the image repair model is obtained by training by using the method described in the first aspect.
  • an apparatus for training an image repair model includes:
  • a generating module which is used to perform damage processing on a complete image by using a preset mask image to generate a damaged image
  • a repair module configured to input the damaged image into an image repair model to obtain a repair image and a missing image, where the missing image is an image of a missing portion of the damaged image;
  • a determining module configured to determine a loss value of the image repair model according to the complete image, the repair image, the mask image, and the missing image;
  • An update module is configured to update the image repair model according to the loss value when the loss value is greater than a preset threshold.
  • an image repair apparatus including:
  • a receiving module for receiving an image to be repaired
  • a processing module is configured to input the image to be repaired into an image repair model to obtain a repaired image, where the image repair model is obtained by training by using the method described in the first aspect.
  • a computer-readable storage medium having stored thereon a computer program that, when executed by a processor, implements the steps of the method described in the first aspect above.
  • a computer-readable storage medium having stored thereon a computer program that, when executed by a processor, implements the steps of the method described in the second aspect above.
  • an electronic device including:
  • a processor configured to execute the computer program in the memory to implement the steps of the method according to the first aspect.
  • an electronic device including:
  • a processor configured to execute the computer program in the memory to implement the steps of the method according to the second aspect.
  • the complete image is damaged by a preset mask image to generate a broken image; and the loss value of the image repair model is determined according to the complete image, the repair image, the mask image, and the missing image, so as to pass This loss value is used to train the image repair model. Therefore, through the above technical solution, when training an image repair model, not only supervised training is performed on the complete image, but also supervised training is performed on the missing image, so that when performing image repair based on the image repair model, the damaged image can be effectively guaranteed The accuracy of the missing part can effectively improve the accuracy of the image repair model. Furthermore, the robustness and migration of the image repair model can be improved, and the quality of the repaired image obtained by image repair based on the image repair model can be effectively improved, and the user experience can be improved.
  • FIG. 1 is a flowchart of a method for training an image repair model according to an embodiment of the present disclosure
  • FIG. 2 is a flowchart of an exemplary implementation manner for determining a loss value of an image repair model based on a complete image, a repaired image, a mask image, and a missing image;
  • FIG. 3 is a flowchart of an image repair method according to an embodiment of the present disclosure.
  • FIG. 4 is a block diagram of a training image repair model provided according to an embodiment of the present disclosure.
  • FIG. 5 is a block diagram of a determination module for a training image repair model provided according to another embodiment of the present disclosure.
  • FIG. 6 is a block diagram of an image repair apparatus according to another embodiment of the present disclosure.
  • Fig. 7 is a block diagram of an electronic device according to an exemplary embodiment
  • Fig. 8 is a block diagram of an electronic device according to an exemplary embodiment.
  • FIG. 1 is a flowchart of a method for training an image repair model according to an embodiment of the present disclosure. As shown in FIG. 1, the method includes:
  • damage processing is performed on the complete image by using a preset mask image to generate a damaged image.
  • the preset image is used to perform damage processing on the complete image to generate a damaged image, and at the same time, the correspondence between the mask image and the damaged image can be stored.
  • the preset mask image may be various noises randomly generated.
  • the damaged image is input to an image repair model to obtain a repaired image and a missing image, where the missing image is an image of a missing portion of the damaged image.
  • the image repair model can be implemented by a self-encoding full convolutional network of any structure. For example, the last feature layer in the convolutional network learns the missing image corresponding to the damaged image, where the missing image is the same size as the damaged image.
  • the repair image is obtained by superimposing the damaged image and the missing image.
  • the repaired image can be determined by the following formula:
  • H (x) represents an image matrix of the missing image
  • x represents an image matrix of the damaged image.
  • the last feature layer of the image repair model may output a missing image corresponding to the input damaged image, and both have the same size. Therefore, the damaged image and the missing image can be superimposed to obtain a repaired image. Therefore, through the above technical solution, when training an image repair model, on the one hand, it can learn to train the missing images corresponding to the broken image, improving the applicability and transferability of the image repair model, and on the other hand, it can quickly determine Repair the image.
  • the loss value of the image repair model is determined according to the complete image, the repaired image, the mask image, and the missing image.
  • the size of the loss value can represent the quality of the repaired image obtained by the image repair model. When the loss value is smaller, the quality of the repaired image is higher. Therefore, in determining the loss value of the image repair model, as compared with only the loss between the complete image and the repaired image in the prior art, the present disclosure determines the image repair through the complete image, repaired image, mask image, and missing image
  • the loss value of the model enables the image repair model to be supervised and trained through multiple images to broaden the scope of application of the image repair model.
  • the preset threshold may represent an acceptance degree of image repair loss of the image repair model.
  • the preset threshold may be 10 -6 . Therefore, when the loss value is greater than the preset threshold of 10-6 , it means that the loss value of the image repair model is large at this time, that is, the training needs to be continued to obtain a more accurate image repair model. It should be noted that the method for updating the image repair model according to the loss value is the prior art, and details are not described herein again.
  • the main purpose is to restore the missing parts in the damaged image.
  • the complete image is damaged by a preset mask image to generate a broken image; and the loss value of the image repair model is determined according to the complete image, the repair image, the mask image, and the missing image, so as to pass This loss value is used to train the image repair model. Therefore, through the above technical solution, when training an image repair model, not only supervised training is performed on the complete image, but also supervised training is performed on the missing image, so that when performing image repair based on the image repair model, the damaged image can be effectively guaranteed. The accuracy of the missing part can effectively improve the accuracy of the image repair model. Furthermore, the robustness and migration of the image repair model can be improved, and the quality of the repaired image obtained by image repair based on the image repair model can be effectively improved, and the user experience can be improved.
  • the updated image repair model after updating the image repair model, return to the step of performing a damage process on the complete image by using a preset mask image to generate a damaged image, until the updated image repair model has a loss value less than Or equal to the preset threshold.
  • the new damaged image generated may be obtained by performing damage processing on another new complete image according to the mask image used in the previous training. Then, the image repair model is trained by using the mask image, the new complete image, and the new damaged image. The subsequent training steps have been described in detail above, and are not repeated here.
  • the new damaged image generated may be obtained by performing a damage process on the complete image used in the previous training according to another new mask image. Then, the image repair model is trained by the complete image, a new mask image, and a new damaged image. The subsequent training steps have been described in detail above, and are not repeated here.
  • the new damaged image generated may be obtained by performing damage processing on another new complete image according to another new mask image. After that, the image repair model is trained through the new complete image, new mask image, and new damaged image. The subsequent training steps have been described in detail above, and are not repeated here.
  • the above steps (S11-S14) are executed cyclically, thereby implementing the training of the image repair model.
  • the loss value of the updated image repair model is less than or equal to a preset threshold, it indicates that the loss value of the image repair model has satisfied the acceptance degree of the image repair loss, that is, the training process of the image repair model can be ended at this time. After that, image repair can be performed on the damaged image based on the image repair model. Therefore, through the above technical solution, the image repair model can be trained quickly and accurately, thereby reducing the loss value of the image repair model, improving the accuracy and training efficiency of the image repair model, and improving the user experience.
  • step S13 of determining the loss value of the image repair model is as follows, as shown in FIG. 2, which may include:
  • a first loss value is determined based on the complete image and the repaired image.
  • the first loss value may be an Euclidean distance between the image matrix of the complete image and the image matrix of the repaired image.
  • the first loss value can be determined by the following formula:
  • loss1 represents the first loss value
  • Y represents an image matrix of the complete image.
  • a second loss value is determined based on the missing image and the mask image.
  • the second loss value may be a second norm of a matrix obtained by multiplying elements in the image matrix of the missing image and elements in corresponding positions in the image matrix of the mask image.
  • the second loss value can be determined by the following formula:
  • loss2 represents the second loss value
  • M (x) represents the image matrix of the mask image
  • H (x) represents the image matrix of the missing image.
  • a loss value of the image repair model is determined based on the first loss value and the second loss value.
  • an average value of the first loss value and the second loss value may be determined as a loss value of the image repair model.
  • determining the loss value of the image repair model according to the first loss value and the second loss value may include:
  • the loss value of the image repair model is determined by the following formula:
  • loss1 represents the first loss value
  • loss2 represents the second loss value
  • ⁇ represents a preset weight coefficient, and the value range is [0,1], where ⁇ can be determined through experience or experiment.
  • the first loss value is determined by the complete image and the repair image, and the difference loss between the repaired image and the complete image obtained by the image repair model can be obtained; the second loss value is determined by the missing image and the mask image , We can get the difference between the missing image and the mask image obtained by the image repair model.
  • the first loss value between the complete image and the repaired image is taken as the main loss value, so that the repaired image obtained by the image repair model is closer to the complete image.
  • the effect of the second loss value between the missing image and the mask image on the final repair result is also fully considered, so that the image repair model can be applied to the repair of more complex broken images, and at the same time, the image repair model is obtained.
  • the repaired image is more accurate.
  • the missing and repaired images determined by the image repair model can be fully supervised and trained, thereby effectively improving the image repair model.
  • the training efficiency of the image repair model is reduced, and the accuracy of the image repair model is improved.
  • the present disclosure also provides an image repair method. As shown in FIG. 3, the method includes:
  • the image to be repaired is input to an image repair model to obtain a repaired image, wherein the image repair model is obtained by training by using the method for training an image repair model described above.
  • the image repair model when the image repair model is trained, it not only corrects the loss between the complete image and the repaired image, but also corrects the loss between the missing image in the damaged image and the corresponding mask image. Therefore, When the image to be repaired is repaired based on the image repair model, the determined missing image is more fitted to the damaged part in the image to be repaired, thereby improving the quality of the repaired image. In addition, it can also effectively improve the efficiency of image repair and improve the user experience.
  • FIG. 4 is a block diagram of an apparatus for training an image repair model according to an embodiment of the present disclosure. As shown in FIG. 4, the apparatus 10 includes:
  • a generating module 100 configured to perform damage processing on a complete image by using a preset mask image to generate a damaged image
  • a repairing module 200 configured to input the damaged image into an image repair model to obtain a repaired image and a missing image, where the missing image is an image of a missing portion of the damaged image;
  • a determining module 300 configured to determine a loss value of the image repair model according to the complete image, the repair image, the mask image, and the missing image;
  • An update module 400 is configured to update the image repair model according to the loss value when the loss value is greater than a preset threshold.
  • the generation module 100 is triggered to perform a damage process on the complete image by using a preset mask image to generate a damaged image until the updated image is repaired
  • the loss value of the model is less than or equal to the preset threshold.
  • the repair image is obtained by superimposing the damaged image and the missing image.
  • the determining module 300 may include:
  • a first determining submodule 301 configured to determine a first loss value according to the complete image and the repaired image
  • a second determining submodule 302 configured to determine a second loss value according to the missing image and the mask image
  • a third determining sub-module 303 is configured to determine a loss value of the image repair model according to the first loss value and the second loss value.
  • the first loss value is an Euclidean distance between an image matrix of the complete image and an image matrix of the repaired image.
  • the second loss value is a second norm of a matrix obtained by multiplying elements in the image matrix of the missing image and elements in corresponding positions in the image matrix of the mask image.
  • the third determining sub-module 303 determines a loss value of the image repair model by using the following formula:
  • loss1 represents the first loss value
  • loss2 represents the second loss value
  • ⁇ represents a preset weight coefficient
  • the apparatus 20 includes:
  • a receiving module 500 configured to receive an image to be repaired
  • the processing module 600 is configured to input the image to be repaired into an image repair model to obtain a repaired image, where the image repair model is obtained by training by using the method for training an image repair model.
  • Fig. 7 is a block diagram of an electronic device 700 according to an exemplary embodiment.
  • the electronic device 700 may include a processor 701 and a memory 702.
  • the electronic device 700 may further include one or more of a multimedia component 703, an input / output (I / O) interface 704, and a communication component 705.
  • the processor 701 is configured to control the overall operation of the electronic device 700 to complete all or part of the above-mentioned method for training an image repair model or an image repair method.
  • the memory 702 is used to store various types of data to support the operation on the electronic device 700. These data may include, for example, instructions for any application program or method for operating on the electronic device 700, and application-related data. Examples include contact data, messages sent and received, pictures, audio, video, and more.
  • the memory 702 may be implemented by any type of volatile or non-volatile memory device or a combination thereof, such as Static Random Access Memory (SRAM), electrically erasable programmable read-only memory (SRAM) Electrically Erasable Programmable Read-Only Memory (referred to as EEPROM), Erasable Programmable Read-Only Memory (referred to as EPROM), Programmable Read-Only Memory (referred to as PROM), read-only Memory (Read-Only Memory, ROM for short), magnetic memory, flash memory, magnetic disk or optical disk.
  • the multimedia component 703 may include a screen and an audio component. The screen may be, for example, a touch screen, and the audio component is used to output and / or input audio signals.
  • the audio component may include a microphone for receiving external audio signals.
  • the received audio signal may be further stored in the memory 702 or transmitted through the communication component 705.
  • the audio component also includes at least one speaker for outputting an audio signal.
  • the I / O interface 704 provides an interface between the processor 701 and other interface modules.
  • the other interface modules may be a keyboard, a mouse, a button, and the like. These buttons can be virtual buttons or physical buttons.
  • the communication component 705 is used for wired or wireless communication between the electronic device 700 and other devices. Wireless communication, such as Wi-Fi, Bluetooth, Near Field Communication (NFC), 2G, 3G, or 4G, or one or more of them, so the corresponding communication component 705 may include: Wi-Fi module, Bluetooth module, NFC module.
  • the electronic device 700 may be implemented by one or more Application Specific Integrated Circuits (ASICs), digital signal processors (DSPs), and digital signal processing devices (Digital Signal Processing Device (DSPD for short), Programmable Logic Device (PLD), Field Programmable Gate Array (FPGA), controller, microcontroller, microprocessor or other electronic components Implementation, for performing the above-mentioned method of training an image repair model or an image repair method.
  • ASICs Application Specific Integrated Circuits
  • DSPs digital signal processors
  • DSPD Digital Signal Processing Device
  • PLD Programmable Logic Device
  • FPGA Field Programmable Gate Array
  • controller microcontroller, microprocessor or other electronic components Implementation, for performing the above-mentioned method of training an image repair model or an image repair method.
  • a computer-readable storage medium including program instructions is provided.
  • the program instructions when executed by a processor, implement the steps of the above-mentioned method for training an image repair model or an image repair method.
  • the computer-readable storage medium may be the above-mentioned memory 702 including program instructions, and the above-mentioned program instructions may be executed by the processor 701 of the electronic device 700 to complete the above-mentioned method for training an image repair model or an image repair method.
  • Fig. 8 is a block diagram of an electronic device 800 according to an exemplary embodiment.
  • the electronic device 800 may be provided as a server.
  • the electronic device 800 includes processors 822, the number of which may be one or more, and a memory 832 for storing a computer program executable by the processor 822.
  • the computer program stored in the memory 832 may include one or more modules each corresponding to a set of instructions.
  • the processor 822 may be configured to execute the computer program to perform the above-mentioned method of training an image repair model or an image repair method.
  • the electronic device 800 may further include a power supply component 826 and a communication component 850.
  • the power supply component 826 may be configured to perform power management of the electronic device 800.
  • the communication component 850 may be configured to implement communication of the electronic device 800, for example, wired Or wireless communication.
  • the electronic device 800 may further include an input / output (I / O) interface 858.
  • the electronic device 800 can operate based on an operating system stored in the memory 832, such as Windows ServerTM, Mac OSXTM, UnixTM, LinuxTM, and the like.
  • a computer-readable storage medium including program instructions is provided.
  • the program instructions when executed by a processor, implement the steps of the above-mentioned method for training an image repair model or an image repair method.
  • the computer-readable storage medium may be the above-mentioned memory 832 including program instructions, and the above-mentioned program instructions may be executed by the processor 822 of the electronic device 800 to complete the above-mentioned method for training an image repair model or an image repair method.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

一种训练图像修复模型的方法、图像修复方法、装置、介质及设备，其中，所述训练图像修复模型的方法，包括：利用预设的掩膜图像对完整图像进行破损处理，生成破损图像(S11)；将破损图像输入至图像修复模型，获得修复图像和缺失图像(S12)；根据完整图像、修复图像、掩膜图像和缺失图像，确定图像修复模型的损失值(S13)；在损失值大于预设阈值时，根据损失值更新图像修复模型(S14)。因此，可以提高该图像修复模型的鲁棒性和可迁移性，并且有效提高基于该图像修复模型进行图像修复获得的修复图像的质量，提升用户使用体验。

Description

训练图像修复模型的方法、图像修复方法、装置、介质及设备 技术领域
本公开涉及图像处理领域，具体地，涉及一种训练图像修复模型的方法、图像修复方法、装置、介质及设备。
背景技术
图像修复的主要目的是还原图像中缺失的部分，在现实生活中有着广泛的应用。现有技术中，图像修复通常通过深度学习方法实现，如通过训练自编码全卷积神经网络或者对抗生成网络获得图像修复模型，从而基于该图像修复模型获得修复的图像。然而在该图像修复方式中，只通过计算上述网络修复的图像与完整图像之间的像素差异性来监督训练，使得训练所得的图像修复模型鲁棒性较低。
发明内容
本公开的目的是提供一种应用广泛的训练图像修复模型的方法、图像修复方法、装置、介质及设备。
为了实现上述目的，根据本公开的第一方面，提供一种训练图像修复模型的方法，所述方法包括：
利用预设的掩膜图像对完整图像进行破损处理，生成破损图像；
将所述破损图像输入至图像修复模型，获得修复图像和缺失图像，所述缺失图像为所述破损图像的缺失部分的图像；
根据所述完整图像、所述修复图像、所述掩膜图像和所述缺失图像，确定所述图像修复模型的损失值；
在所述损失值大于预设阈值时，根据所述损失值更新所述图像修复模 型。
根据本公开的第二方面，提供一种图像修复方法，所述方法包括：
接收待修复图像；
将所述待修复图像输入至图像修复模型，得到修复图像，其中，所述图像修复模型是通过上述第一方面所述方法进行训练得到的。
根据本公开的第三方面，提供一种训练图像修复模型的装置，所述装置包括：
生成模块，用于利用预设的掩膜图像对完整图像进行破损处理，生成破损图像；
修复模块，用于将所述破损图像输入至图像修复模型，获得修复图像和缺失图像，所述缺失图像为所述破损图像的缺失部分的图像；
确定模块，用于根据所述完整图像、所述修复图像、所述掩膜图像和所述缺失图像，确定所述图像修复模型的损失值；
更新模块，用于在所述损失值大于预设阈值时，根据所述损失值更新所述图像修复模型。
根据本公开的第四方面，提供一种图像修复装置，所述装置包括：
接收模块，用于接收待修复图像；
处理模块，用于将所述待修复图像输入至图像修复模型，得到修复图像，其中，所述图像修复模型是通过上述第一方面所述方法进行训练得到的。
根据本公开的第五方面，提供一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现上述第一方面所述方法的步骤。
根据本公开的第六方面，提供一种计算机可读存储介质，其上存储有计算机程序，该程序被处理器执行时实现上述第二方面所述方法的步骤。
根据本公开的第七方面，提供一种电子设备，包括：
存储器，其上存储有计算机程序；
处理器，用于执行所述存储器中的所述计算机程序，以实现上述第一方面所述方法的步骤。
根据本公开的第八方面，提供一种电子设备，包括：
存储器，其上存储有计算机程序；
处理器，用于执行所述存储器中的所述计算机程序，以实现上述第二方面所述方法的步骤。
在上述技术方案中，通过预设的掩膜图像对完整图像进行破损处理，以生成破损图像；并根据完整图像、修复图像、掩膜图像、缺失图像，确定图像修复模型的损失值，从而通过该损失值对图像修复模型进行训练。因此，通过上述技术方案，在对图像修复模型进行训练时，不仅通过完整图像进行监督训练，同时还通过缺失图像进行监督训练，使得在基于图像修复模型进行图像修复时，可以有效保证破损图像中的缺失部分的准确性，从而可以有效提高图像修复模型的精准度。更进一步地，还可以提高该图像修复模型的鲁棒性和可迁移性，并且有效提高基于该图像修复模型进行图像修复获得的修复图像的质量，提升用户使用体验。
本公开的其他特征和优点将在随后的具体实施方式部分予以详细说明。
附图说明
附图是用来提供对本公开的进一步理解，并且构成说明书的一部分，与下面的具体实施方式一起用于解释本公开，但并不构成对本公开的限制。在附图中：
图1是根据本公开的一种实施方式提供的训练图像修复模型的方法的流程图；
图2是根据完整图像、修复图像、掩膜图像和缺失图像，确定图像修复模型的损失值的一种示例性实现方式的流程图；
图3是根据本公开的一种实施方式提供的一种图像修复方法的流程图；
图4是根据本公开的一种实施方式提供的训练图像修复模型的框图；
图5是根据本公开的另一种实施方式提供的训练图像修复模型的确定模块的框图；
图6是根据本公开的另一种实施方式提供的图像修复装置的框图；
图7是根据一示例性实施例示出的一种电子设备的框图；
图8是根据一示例性实施例示出的一种电子设备的框图。
具体实施方式
以下结合附图对本公开的具体实施方式进行详细说明。应当理解的是，此处所描述的具体实施方式仅用于说明和解释本公开，并不用于限制本公开。
图1所示，为根据本公开的一种实施方式提供的训练图像修复模型的方法的流程图。如图1所示，所述方法包括：
在S11中，利用预设的掩膜图像对完整图像进行破损处理，生成破损图像。
其中，可以从数据库或者互联网上获取大量的完整图像。之后，利用预设的掩膜图像对完整图像进行破损处理，生成破损图像，同时可以存储该掩膜图像与破损图像之间的对应关系。其中，预设的掩膜图像可以是随机生成的各种噪声。
在S12中，将破损图像输入至图像修复模型，获得修复图像和缺失图像，所述缺失图像为所述破损图像的缺失部分的图像。
其中，所述图像修复模型可以通过任意结构的自编码全卷积网络实现。示例地，在该卷积网络中的最后一个特征层学习的是破损图像对应的缺失图像，其中，缺失图像与破损图像的大小相同。
可选地，所述修复图像通过如下方式获得：将所述破损图像与所述缺失图像叠加。
示例地，可以通过如下公式确定修复图像：
Figure PCTCN2018094900-appb-000001
其中，
Figure PCTCN2018094900-appb-000002
表示所述修复图像的图像矩阵；
H(x)表示所述缺失图像的图像矩阵；
x表示所述破损图像的图像矩阵。
在该实施例中，当破损图像输入至图像修复模型中时，图像修复模型的最后一个特征层可以输出与输入的破损图像对应的缺失图像，并且两者大小相同。因此，可以将该破损图像和缺失图像进行叠加，从而获得修复图像。因此，通过上述技术方案，在对图像修复模型进行训练时，一方面可以学习训练破损图像对应的缺失图像，提高该图像修复模型的适用性和可迁移性，另一方面，又可以快速确定出修复图像。
在S13中，根据完整图像、修复图像、掩膜图像和缺失图像，确定图像修复模型的损失值。
其中，所述损失值的大小可以表征通过图像修复模型获得的修复图像的质量，当损失值越小时，表征修复图像的质量越高。因此，在确定图像修复模型的损失值时，与现有技术中只考虑完整图像与修复图像之间的损失相较，本公开中通过完整图像、修复图像、掩膜图像和缺失图像确定图像修复模型的损失值，使得图像修复模型通过多个图像进行监督训练，以拓宽图像修复模型的适用范围。
在S14中，在损失值大于预设阈值时，根据损失值更新图像修复模型。
其中，预设阈值可以表示该图像修复模型的图像修复损失的接受程度，示例地，该预设阈值可以为10 -6。因此，当损失值大于该预设阈值10 -6时，表示此时图像修复模型的损失值较大，即需要继续进行训练以获得更加准确的 图像修复模型。需要进行说明的是，根据损失值更新图像修复模型的方法为现有技术，对此不再赘述。
在对破损图像进行修复时，主要目的是要还原破损图像中的缺失部分。在上述技术方案中，通过预设的掩膜图像对完整图像进行破损处理，以生成破损图像；并根据完整图像、修复图像、掩膜图像、缺失图像，确定图像修复模型的损失值，从而通过该损失值对图像修复模型进行训练。因此，通过上述技术方案，在对图像修复模型进行训练时，不仅通过完整图像进行监督训练，同时还通过缺失图像进行监督训练，使得在基于图像修复模型进行图像修复时，可以有效保证破损图像中的缺失部分的准确性，从而可以有效提高图像修复模型的精准度。更进一步地，还可以提高该图像修复模型的鲁棒性和可迁移性，并且有效提高基于该图像修复模型进行图像修复获得的修复图像的质量，提升用户使用体验。
可选地，在更新所述图像修复模型之后，返回所述利用预设的掩膜图像对完整图像进行破损处理，生成破损图像的步骤S11，直到更新后的所述图像修复模型的损失值小于或等于所述预设阈值为止。
其中，所述预设的掩膜图像、完整图像等可以为多个。在一实施例中，在返回步骤S11时，生成的新的破损图像可以是根据上一训练所用的掩膜图像对另一新的完整图像进行破损处理所得的。之后，通过该掩膜图像、新的完整图像、新的破损图像对该图像修复模型进行训练。其中，后续训练的步骤在上文已经详述，在此不再赘述。
在另一实施例中，在返回步骤S11时，生成的新的破损图像可以是根据另一新的掩膜图像对上一训练所用的完整图像进行破损处理所得的。之后，通过该完整图像、新的掩膜图像、新的破损图像对该图像修复模型进行训练。其中，后续训练的步骤在上文已经详述，在此不再赘述。
在另一实施例中，在返回步骤S11时，生成的新的破损图像可以是根据 另一新的掩膜图像对另一新的完整图像进行破损处理所得的。之后，通过该新的完整图像、新的掩膜图像、新的破损图像对该图像修复模型进行训练。其中，后续训练的步骤在上文已经详述，在此不再赘述。
在上述技术方案中，通过循环执行上述步骤(S11-S14)，从而实现对图像修复模型的训练。当更新后的图像修复模型的损失值小于或等于预设阈值时，表示图像修复模型的损失值已经满足图像修复损失的接受程度，即此时可以结束该图像修复模型的训练过程。之后，可以基于该图像修复模型对破损图像进行图像修复。因此，通过上述技术方案，可以快速且准确地对图像修复模型进行训练，从而降低图像修复模型的损失值，提高图像修复模型的精准度和训练效率，提升用户使用体验。
可选地，根据完整图像、修复图像、掩膜图像和缺失图像，确定图像修复模型的损失值的步骤S13的一种示例性实现方式如下，如图2所示，可以包括：
在S21中，根据完整图像和修复图像，确定第一损失值。
其中，所述第一损失值可以为所述完整图像的图像矩阵与所述修复图像的图像矩阵之间的欧氏距离。示例地，可以通过如下公式确定第一损失值：
Figure PCTCN2018094900-appb-000003
其中，loss1表示所述第一损失值；
Figure PCTCN2018094900-appb-000004
表示所述修复图像的图像矩阵；
Y表示所述完整图像的图像矩阵。
在S22中，根据缺失图像和掩膜图像，确定第二损失值。
其中，所述第二损失值可以为将所述缺失图像的图像矩阵中的元素与所述掩膜图像的图像矩阵中对应位置的元素相乘所得的矩阵的二范数。示例地，可以通过如下公式确定第二损失值：
loss2＝||M(x)⊙H(x)|| 2；
其中，loss2表示所述第二损失值；
M(x)表示掩膜图像的图像矩阵；
H(x)表示缺失图像的图像矩阵。
在S23中，根据第一损失值和第二损失值，确定图像修复模型的损失值。
在一实施例中，可以将所述第一损失值和所述第二损失值的平均值确定为所述图像修复模型的损失值。
在另一实施例中，所述根据所述第一损失值和所述第二损失值，确定所述图像修复模型的损失值，可以包括：
通过如下公式确定所述图像修复模型的损失值：
loss＝loss1+λloss2
其中，loss1表示所述第一损失值；
loss2表示所述第二损失值；
λ表示预设的权重系数，取值范围为[0,1]，其中，λ可以通过经验或者试验确定。
在上述技术方案中，通过完整图像和修复图像确定第一损失值，可以得出通过图像修复模型获得的修复图像和完整图像之间的差异损失；通过缺失图像和掩膜图像确定第二损失值，可以得出通过图像修复模型获得的缺失图像和掩膜图像之间的差异损失。在通过上述公式确定图像修复模型的损失值时，将完整图像和修复图像之间的第一损失值作为主要损失值，使得通过该图像修复模型的获得的修复图像更加接近于完整图像。同时，还充分考虑了缺失图像和掩膜图像之间的第二损失值对最终修复结果的影响，使得该图像修复模型可以适用于更加复杂的破损图像的修复，同时使得通过图像修复模型的获得的修复图像更加精准。
因此，通过第一损失值和第二损失值确定出的损失值对图像修复模型进行监督训练时，可以充分对图像修复模型确定出的缺失图像和修复图像进行 监督训练，从而有效提高图像修复模型的训练效率，降低图像修复模型的训练数量，提高图像修复模型的图像修复的精准度。
本公开还提供一种图像修复方法，如图3所示，所述方法包括：
在S31中，接收待修复图像；
在S32中，将待修复图像输入至图像修复模型，得到修复图像，其中，所述图像修复模型是通过上述的训练图像修复模型的方法进行训练得到的。
在该实施例中，图像修复模型在训练时，不仅对完整图像与修复图像之间的损失进行修正，还对破损图像中的缺失图像与对应的掩膜图像之间的损失进行修正，因此，在基于图像修复模型对待修复图像进行修复时，使得确定出的缺失图像更加贴合待修复图像中的破损部分，从而提高修复图像的质量。另外，也可以有效提高图像修复的效率，提升用户使用体验。
本公开还提供一种训练图像修复模型的装置。图4所示，为根据本公开的一种实施方式提供的训练图像修复模型的装置的框图，如图4所示，所述装置10包括：
生成模块100，用于利用预设的掩膜图像对完整图像进行破损处理，生成破损图像；
修复模块200，用于将所述破损图像输入至图像修复模型，获得修复图像和缺失图像，所述缺失图像为所述破损图像的缺失部分的图像；
确定模块300，用于根据所述完整图像、所述修复图像、所述掩膜图像和所述缺失图像，确定所述图像修复模型的损失值；
更新模块400，用于在所述损失值大于预设阈值时，根据所述损失值更新所述图像修复模型。
可选地，在所述更新模块400更新所述图像修复模型之后，触发所述生成模块100利用预设的掩膜图像对完整图像进行破损处理，生成破损图像，直到更新后的所述图像修复模型的损失值小于或等于所述预设阈值为止。
可选地，所述修复图像通过如下方式获得：将所述破损图像与所述缺失图像叠加。
可选地，如图5所示，所述确定模块300可以包括：
第一确定子模块301，用于根据所述完整图像和所述修复图像，确定第一损失值；
第二确定子模块302，用于根据所述缺失图像和所述掩膜图像，确定第二损失值；
第三确定子模块303，用于根据所述第一损失值和所述第二损失值，确定所述图像修复模型的损失值。
可选地，所述第一损失值为所述完整图像的图像矩阵与所述修复图像的图像矩阵之间的欧氏距离。
可选地，所述第二损失值为将所述缺失图像的图像矩阵中的元素与所述掩膜图像的图像矩阵中对应位置的元素相乘所得的矩阵的二范数。
可选地，所述第三确定子模块303通过如下公式确定所述图像修复模型的损失值：
loss＝loss1+λloss2
其中，loss1表示所述第一损失值；
loss2表示所述第二损失值；
λ表示预设的权重系数。
本公开还提供一种图像修复装置，如图6所示，所述装置20包括：
接收模块500，用于接收待修复图像；
处理模块600，用于将所述待修复图像输入至图像修复模型，得到修复图像，其中，所述图像修复模型是通过上述训练图像修复模型的方法进行训练得到的。
关于上述实施例中的装置，其中各个模块执行操作的具体方式已经在有 关该方法的实施例中进行了详细描述，此处将不做详细阐述说明。
图7是根据一示例性实施例示出的一种电子设备700的框图。如图7所示，该电子设备700可以包括：处理器701，存储器702。该电子设备700还可以包括多媒体组件703，输入/输出(I/O)接口704，以及通信组件705中的一者或多者。
其中，处理器701用于控制该电子设备700的整体操作，以完成上述的训练图像修复模型的方法或者图像修复方法中的全部或部分步骤。存储器702用于存储各种类型的数据以支持在该电子设备700的操作，这些数据例如可以包括用于在该电子设备700上操作的任何应用程序或方法的指令，以及应用程序相关的数据，例如联系人数据、收发的消息、图片、音频、视频等等。该存储器702可以由任何类型的易失性或非易失性存储设备或者它们的组合实现，例如静态随机存取存储器(Static Random Access Memory，简称SRAM)，电可擦除可编程只读存储器(Electrically Erasable Programmable Read-Only Memory，简称EEPROM)，可擦除可编程只读存储器(Erasable Programmable Read-Only Memory，简称EPROM)，可编程只读存储器(Programmable Read-Only Memory，简称PROM)，只读存储器(Read-Only Memory，简称ROM)，磁存储器，快闪存储器，磁盘或光盘。多媒体组件703可以包括屏幕和音频组件。其中屏幕例如可以是触摸屏，音频组件用于输出和/或输入音频信号。例如，音频组件可以包括一个麦克风，麦克风用于接收外部音频信号。所接收的音频信号可以被进一步存储在存储器702或通过通信组件705发送。音频组件还包括至少一个扬声器，用于输出音频信号。I/O接口704为处理器701和其他接口模块之间提供接口，上述其他接口模块可以是键盘，鼠标，按钮等。这些按钮可以是虚拟按钮或者实体按钮。通信组件705用于该电子设备700与其他设备之间进行有线或无线通信。无线通信，例如Wi-Fi，蓝牙，近场通信(Near Field Communication，简称NFC)， 2G、3G或4G，或它们中的一种或几种的组合，因此相应的该通信组件705可以包括：Wi-Fi模块，蓝牙模块，NFC模块。
在一示例性实施例中，电子设备700可以被一个或多个应用专用集成电路(Application Specific Integrated Circuit，简称ASIC)、数字信号处理器(Digital Signal Processor，简称DSP)、数字信号处理设备(Digital Signal Processing Device，简称DSPD)、可编程逻辑器件(Programmable Logic Device，简称PLD)、现场可编程门阵列(Field Programmable Gate Array，简称FPGA)、控制器、微控制器、微处理器或其他电子元件实现，用于执行上述的训练图像修复模型的方法或者图像修复方法。
在另一示例性实施例中，还提供了一种包括程序指令的计算机可读存储介质，该程序指令被处理器执行时实现上述的训练图像修复模型的方法或者图像修复方法的步骤。例如，该计算机可读存储介质可以为上述包括程序指令的存储器702，上述程序指令可由电子设备700的处理器701执行以完成上述的训练图像修复模型的方法或者图像修复方法。
图8是根据一示例性实施例示出的一种电子设备800的框图。例如，电子设备800可以被提供为一服务器。参照图8，电子设备800包括处理器822，其数量可以为一个或多个，以及存储器832，用于存储可由处理器822执行的计算机程序。存储器832中存储的计算机程序可以包括一个或一个以上的每一个对应于一组指令的模块。此外，处理器822可以被配置为执行该计算机程序，以执行上述的训练图像修复模型的方法或者图像修复方法。
另外，电子设备800还可以包括电源组件826和通信组件850，该电源组件826可以被配置为执行电子设备800的电源管理，该通信组件850可以被配置为实现电子设备800的通信，例如，有线或无线通信。此外，该电子设备800还可以包括输入/输出(I/O)接口858。电子设备800可以操作基于存储在存储器832的操作***，例如Windows ServerTM，Mac OS XTM， UnixTM,LinuxTM等等。
在另一示例性实施例中，还提供了一种包括程序指令的计算机可读存储介质，该程序指令被处理器执行时实现上述的训练图像修复模型的方法或者图像修复方法的步骤。例如，该计算机可读存储介质可以为上述包括程序指令的存储器832，上述程序指令可由电子设备800的处理器822执行以完成上述的训练图像修复模型的方法或者图像修复方法。
以上结合附图详细描述了本公开的优选实施方式，但是，本公开并不限于上述实施方式中的具体细节，在本公开的技术构思范围内，可以对本公开的技术方案进行多种简单变型，这些简单变型均属于本公开的保护范围。
另外需要说明的是，在上述具体实施方式中所描述的各个具体技术特征，在不矛盾的情况下，可以通过任何合适的方式进行组合。为了避免不必要的重复，本公开对各种可能的组合方式不再另行说明。
此外，本公开的各种不同的实施方式之间也可以进行任意组合，只要其不违背本公开的思想，其同样应当视为本公开所公开的内容。

Claims (14)

  1. 一种训练图像修复模型的方法，其特征在于，所述方法包括：
    利用预设的掩膜图像对完整图像进行破损处理，生成破损图像；
    将所述破损图像输入至图像修复模型，获得修复图像和缺失图像，所述缺失图像为所述破损图像的缺失部分的图像；
    根据所述完整图像、所述修复图像、所述掩膜图像和所述缺失图像，确定所述图像修复模型的损失值；
    在所述损失值大于预设阈值时，根据所述损失值更新所述图像修复模型。
  2. 根据权利要求1所述的方法，其特征在于，在更新所述图像修复模型之后，返回所述利用预设的掩膜图像对完整图像进行破损处理，生成破损图像的步骤，直到更新后的所述图像修复模型的损失值小于或等于所述预设阈值为止。
  3. 根据权利要求1所述的方法，其特征在于，所述修复图像通过如下方式获得：将所述破损图像与所述缺失图像叠加。
  4. 根据权利要求1-3中任一项所述的方法，其特征在于，所述根据所述完整图像、所述修复图像、所述掩膜图像和所述缺失图像，确定所述图像修复模型的损失值，包括：
    根据所述完整图像和所述修复图像，确定第一损失值；
    根据所述缺失图像和所述掩膜图像，确定第二损失值；
    根据所述第一损失值和所述第二损失值，确定所述图像修复模型的损失值。
  5. 根据权利要求4所述的方法，其特征在于，所述第一损失值为所述完整图像的图像矩阵与所述修复图像的图像矩阵之间的欧氏距离。
  6. 根据权利要求4所述的方法，其特征在于，所述第二损失值为将所述缺失图像的图像矩阵中的元素与所述掩膜图像的图像矩阵中对应位置的元素相乘所得的矩阵的二范数。
  7. 根据权利要求4所述的方法，其特征在于，所述根据所述第一损失值和所述第二损失值，确定所述图像修复模型的损失值，包括：
    通过如下公式确定所述图像修复模型的损失值：
    loss＝loss1+λloss2
    其中，loss1表示所述第一损失值；
    loss2表示所述第二损失值；
    λ表示预设的权重系数。
  8. 一种图像修复方法，其特征在于，所述方法包括：
    接收待修复图像；
    将所述待修复图像输入至图像修复模型，得到修复图像，其中，所述图像修复模型是通过权利要求1-7中任一项所述的方法进行训练得到的。
  9. 一种训练图像修复模型的装置，其特征在于，所述装置包括：
    生成模块，用于利用预设的掩膜图像对完整图像进行破损处理，生成破损图像；
    修复模块，用于将所述破损图像输入至图像修复模型，获得修复图像和缺失图像，所述缺失图像为所述破损图像的缺失部分的图像；
    确定模块，用于根据所述完整图像、所述修复图像、所述掩膜图像和所 述缺失图像，确定所述图像修复模型的损失值；
    更新模块，用于在所述损失值大于预设阈值时，根据所述损失值更新所述图像修复模型。
  10. 一种图像修复装置，其特征在于，所述装置包括：
    接收模块，用于接收待修复图像；
    处理模块，用于将所述待修复图像输入至图像修复模型，得到修复图像，其中，所述图像修复模型是通过权利要求1-7中任一项所述的方法进行训练得到的。
  11. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现权利要求1-7中任一项所述方法的步骤。
  12. 一种计算机可读存储介质，其上存储有计算机程序，其特征在于，该程序被处理器执行时实现权利要求8所述方法的步骤。
  13. 一种电子设备，其特征在于，包括：
    存储器，其上存储有计算机程序；
    处理器，用于执行所述存储器中的所述计算机程序，以实现权利要求1-7中任一项所述方法的步骤。
  14. 一种电子设备，其特征在于，包括：
    存储器，其上存储有计算机程序；
    处理器，用于执行所述存储器中的所述计算机程序，以实现权利要求8所述方法的步骤。
PCT/CN2018/094900 2018-07-06 2018-07-06 训练图像修复模型的方法、图像修复方法、装置、介质及设备 WO2020006762A1 (zh)

Priority Applications (2)

Application Number Priority Date Filing Date Title
CN201880001170.8A CN108885787B (zh) 2018-07-06 2018-07-06 训练图像修复模型的方法、图像修复方法、装置、介质及设备
PCT/CN2018/094900 WO2020006762A1 (zh) 2018-07-06 2018-07-06 训练图像修复模型的方法、图像修复方法、装置、介质及设备

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2018/094900 WO2020006762A1 (zh) 2018-07-06 2018-07-06 训练图像修复模型的方法、图像修复方法、装置、介质及设备

Publications (1)

Publication Number Publication Date
WO2020006762A1 true WO2020006762A1 (zh) 2020-01-09

Family

ID=64325021

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2018/094900 WO2020006762A1 (zh) 2018-07-06 2018-07-06 训练图像修复模型的方法、图像修复方法、装置、介质及设备

Country Status (2)

Country Link
CN (1) CN108885787B (zh)
WO (1) WO2020006762A1 (zh)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111462000A (zh) * 2020-03-17 2020-07-28 北京邮电大学 一种基于预训练自编码器的图像恢复方法及装置
CN112418054A (zh) * 2020-11-18 2021-02-26 北京字跳网络技术有限公司 图像处理方法、装置、电子设备和计算机可读介质

Families Citing this family (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110852980A (zh) * 2019-11-13 2020-02-28 北京京东方专用显示科技有限公司 交互式图像填充方法及***、服务器、设备及介质
CN111353956B (zh) * 2020-02-28 2022-11-15 Oppo广东移动通信有限公司 图像修复方法、装置、计算机设备及存储介质
CN111353965B (zh) * 2020-02-28 2023-08-01 Oppo广东移动通信有限公司 图像修复方法、装置、终端及存储介质
CN112614066A (zh) * 2020-12-23 2021-04-06 文思海辉智科科技有限公司 图像修复方法、装置及电子设备
CN113327210B (zh) * 2021-06-30 2023-04-07 中海油田服务股份有限公司 测井图像填补方法、装置、介质及电子设备
CN113744199B (zh) * 2021-08-10 2023-09-26 南方科技大学 图像的破损检测方法、电子设备及存储介质

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102142132A (zh) * 2011-03-31 2011-08-03 北京交通大学 基于模块的图像修复方法
CN103886561A (zh) * 2014-04-09 2014-06-25 武汉科技大学 基于数学形态学的Criminisi图像修复方法
CN104574294A (zh) * 2014-12-12 2015-04-29 中国农业大学 基于Shannon-Nuttall小波多尺度表达的图像修复方法及装置
CN107945140A (zh) * 2017-12-20 2018-04-20 中国科学院深圳先进技术研究院 一种图像修复方法、装置及设备
CN107993210A (zh) * 2017-11-30 2018-05-04 北京小米移动软件有限公司 图像修复方法、装置及计算机可读存储介质

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
KR101341617B1 (ko) * 2012-03-14 2013-12-13 중앙대학교 산학협력단 단일 영상의 오차모델을 기반으로 한 고해상도 영상 복원장치 및 방법
CN105761213B (zh) * 2014-12-16 2019-02-26 北京大学 图像修补方法和图像修补装置
CN106846279B (zh) * 2017-03-02 2019-10-01 合肥工业大学 一种基于连分式插值技术的自适应图像修补方法及其***

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN102142132A (zh) * 2011-03-31 2011-08-03 北京交通大学 基于模块的图像修复方法
CN103886561A (zh) * 2014-04-09 2014-06-25 武汉科技大学 基于数学形态学的Criminisi图像修复方法
CN104574294A (zh) * 2014-12-12 2015-04-29 中国农业大学 基于Shannon-Nuttall小波多尺度表达的图像修复方法及装置
CN107993210A (zh) * 2017-11-30 2018-05-04 北京小米移动软件有限公司 图像修复方法、装置及计算机可读存储介质
CN107945140A (zh) * 2017-12-20 2018-04-20 中国科学院深圳先进技术研究院 一种图像修复方法、装置及设备

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111462000A (zh) * 2020-03-17 2020-07-28 北京邮电大学 一种基于预训练自编码器的图像恢复方法及装置
CN111462000B (zh) * 2020-03-17 2023-11-28 北京邮电大学 一种基于预训练自编码器的图像恢复方法及装置
CN112418054A (zh) * 2020-11-18 2021-02-26 北京字跳网络技术有限公司 图像处理方法、装置、电子设备和计算机可读介质

Also Published As

Publication number Publication date
CN108885787A (zh) 2018-11-23
CN108885787B (zh) 2021-11-05

Similar Documents

Publication Publication Date Title
WO2020006762A1 (zh) 训练图像修复模型的方法、图像修复方法、装置、介质及设备
CN109753910B (zh) 关键点提取方法、模型的训练方法、装置、介质及设备
CN109658346B (zh) 图像修复方法、装置、计算机可读存储介质及电子设备
JP2019528502A (ja) パターン認識に適用可能なモデルを最適化するための方法および装置ならびに端末デバイス
CN111967609B (zh) 模型参数验证方法、设备及可读存储介质
CN111626408B (zh) 哈希编码方法、装置、设备及可读存储介质
CN111105375B (zh) 图像生成方法及其模型训练方法、装置及电子设备
JP2020512639A5 (zh)
CN113642659B (zh) 一种训练样本集生成的方法、装置、电子设备及存储介质
KR20210116923A (ko) 디노이징 네트워크의 학습 방법, 및 이미지 처리 방법 및 그 장치 방법
CN109492531B (zh) 人脸图像关键点提取方法、装置、存储介质及电子设备
CN111625258B (zh) 默克尔树更新方法、装置、设备及可读存储介质
CN109657539B (zh) 人脸颜值评价方法、装置、可读存储介质及电子设备
CN110853663A (zh) 基于人工智能的语音增强方法、服务器及存储介质
JP2019197540A5 (zh)
TWI664853B (zh) 視訊壓縮感測重建方法及裝置
CN114898177B (zh) 缺陷图像生成方法、模型训练方法、设备、介质及产品
CN110751593A (zh) 一种图像虚化处理方法及装置
KR20200049422A (ko) 시뮬레이션-가이드된 반복적 프루닝을 사용하는 효율적인 네트워크 압축
CN113965313B (zh) 基于同态加密的模型训练方法、装置、设备以及存储介质
CN116721179A (zh) 一种基于扩散模型生成图像的方法、设备和存储介质
Liu et al. Learning collaborative generation correction modules for blind image deblurring and beyond
TW202244758A (zh) 指紋識別方法、指紋模組及電子設備
CN107977628B (zh) 神经网络训练方法、人脸检测方法及人脸检测装置
US20220310068A1 (en) Methods and devices for structured pruning for automatic speech recognition

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: 18925239

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

32PN Ep: public notification in the ep bulletin as address of the adressee cannot be established

Free format text: NOTING OF LOSS OF RIGHTS PURSUANT TO RULE 112(1) EPC (EPO FORM 1205A DATED 14.04.2021)

122 Ep: pct application non-entry in european phase

Ref document number: 18925239

Country of ref document: EP

Kind code of ref document: A1