WO2023050868A1 - 融合模型的训练方法、图像融合方法、装置、设备及介质 - Google Patents

融合模型的训练方法、图像融合方法、装置、设备及介质 Download PDF

Info

Publication number
WO2023050868A1
WO2023050868A1 PCT/CN2022/097872 CN2022097872W WO2023050868A1 WO 2023050868 A1 WO2023050868 A1 WO 2023050868A1 CN 2022097872 W CN2022097872 W CN 2022097872W WO 2023050868 A1 WO2023050868 A1 WO 2023050868A1
Authority
WO
WIPO (PCT)
Prior art keywords
training
image
fusion
alignment
attribute
Prior art date
Application number
PCT/CN2022/097872
Other languages
English (en)
French (fr)
Inventor
徐志良
洪智滨
Original Assignee
北京百度网讯科技有限公司
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 北京百度网讯科技有限公司 filed Critical 北京百度网讯科技有限公司
Publication of WO2023050868A1 publication Critical patent/WO2023050868A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/25—Fusion techniques
    • 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 artificial intelligence technology, in particular to the field of computer vision and deep learning technology, which can be applied to scenes such as face image processing and face recognition, and specifically relates to a training method for a fusion model, an image fusion method, a device, an electronic device, and a storage device. media and program products.
  • Image fusion can refer to the technology of combining two or more images into a new image.
  • Image fusion can take advantage of the correlation and complementarity between multiple images, so that the new image obtained after fusion has a more comprehensive and clear content display, which is beneficial to identification and detection. Provide great help for the application development of public security, information security, and financial security.
  • the disclosure provides a fusion model training method, image fusion method, device, electronic equipment, storage medium and program product.
  • a method for training a fusion model including: inputting a training source image and a training template image into the fusion model to obtain a training fusion image; performing attribute alignment transformation on the training fusion image to obtain training the alignment image, wherein the attribute information of the training alignment image is consistent with the attribute information of the training source image; and using an identity loss function to train the fusion model, wherein the identity loss function is for the training source image and the training alignment images are generated.
  • an image fusion method including: inputting the image to be fused and the template image into a fusion model to obtain a fusion image; wherein, the fusion model uses The training method is trained to get.
  • a training device for a fusion model including: a training fusion module for inputting a training source image and a training template image into the fusion model to obtain a training fusion image; an attribute transformation module for Perform attribute alignment transformation on the training fusion image to obtain a training alignment image, wherein the attribute information of the training alignment image is consistent with the attribute information of the training source image; and a training module is used to use the identity loss function to train all The fusion model, wherein the identity loss function is generated for the training source images and the training alignment images.
  • an image fusion device including: a fusion module, configured to input the image to be fused and the template image into the fusion model to obtain a fusion image; wherein, the fusion model utilizes the above-mentioned
  • the training method of the fusion model is obtained by training.
  • an electronic device including: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores Instructions executed by the at least one processor to enable the at least one processor to perform the method as described above.
  • a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause the computer to execute the above method.
  • a computer program product comprising a computer program which, when executed by a processor, implements the method as described above.
  • FIG. 1 schematically shows an exemplary system architecture to which an image fusion method and device can be applied according to an embodiment of the present disclosure
  • Fig. 2 schematically shows a flow chart of a method for training a fusion model according to an embodiment of the present disclosure
  • FIG. 3 schematically shows a schematic flowchart of a fusion model training method according to another embodiment of the present disclosure
  • FIG. 4 schematically shows a schematic flowchart of a training method of a fusion model according to another embodiment of the present disclosure
  • FIG. 5 schematically shows a flowchart of an image fusion method according to an embodiment of the present disclosure
  • FIG. 6 schematically shows a flowchart of an image fusion method according to another embodiment of the present disclosure
  • Fig. 7 schematically shows a block diagram of a training device for a fusion model according to an embodiment of the present disclosure
  • Fig. 8 schematically shows a block diagram of an image fusion device according to an embodiment of the present disclosure.
  • Fig. 9 schematically shows a block diagram of an electronic device suitable for implementing an image fusion method according to an embodiment of the present disclosure.
  • the disclosure provides a fusion model training method, image fusion method, device, electronic equipment, storage medium and program product.
  • the training method of the fusion model may include: inputting the training source image and the training template image into the fusion model to obtain the training fusion image; performing attribute alignment transformation on the training fusion image to obtain the training alignment image, wherein, The attribute information of the training alignment image is consistent with the attribute information of the training source image; and an identity loss function is used to train the fusion model, wherein the identity loss function is generated for the training source image and the training alignment image.
  • the image fusion method includes inputting the image to be fused and the template image into a fusion model to obtain a fusion image, wherein the fusion model is trained by using the fusion model training method provided by the embodiment of the present disclosure.
  • the user's authorization or consent is obtained.
  • Fig. 1 schematically shows an exemplary system architecture to which an image fusion method and device can be applied according to an embodiment of the present disclosure.
  • the exemplary system architecture to which the image fusion method and device can be applied may include a terminal device, but the terminal device may implement the image fusion method and device provided in the embodiments of the present disclosure without interacting with the server .
  • a system architecture 100 may include terminal devices 101 , 102 , 103 , a network 104 and a server 105 .
  • the network 104 is used as a medium for providing communication links between the terminal devices 101 , 102 , 103 and the server 105 .
  • Network 104 may include various connection types, such as wired and/or wireless communication links, among others.
  • Terminal devices 101 , 102 , 103 Users can use terminal devices 101 , 102 , 103 to interact with server 105 through network 104 to send source images and template images and receive fused images.
  • Various communication client applications can be installed on the terminal devices 101, 102, 103, such as an application program loaded with an image fusion method (just an example).
  • the terminal devices 101, 102, 103 may be various electronic devices with display screens and cameras, including but not limited to smartphones, tablet computers, laptop computers, desktop computers, and the like.
  • the server 105 may be a server that provides various services, such as a background management server that provides support for source images and template images uploaded by users using the terminal devices 101 , 102 , 103 (just an example).
  • the background management server can perform image fusion processing on the source image and the template image to obtain a fusion image, and feed back the fusion image to the terminal device.
  • the image fusion method provided by the embodiment of the present disclosure may be executed by the terminal device 101 , 102 , or 103 .
  • the image fusion apparatus provided by the embodiment of the present disclosure may also be set in the terminal device 101 , 102 , or 103 .
  • the image fusion method provided by the embodiment of the present disclosure may also be executed by the server 105 .
  • the image fusion apparatus provided by the embodiment of the present disclosure can generally be set in the server 105 .
  • the image fusion method provided by the embodiments of the present disclosure may also be executed by a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101 , 102 , 103 and/or the server 105 .
  • the image fusion apparatus provided by the embodiments of the present disclosure may also be set in a server or a server cluster that is different from the server 105 and can communicate with the terminal devices 101 , 102 , 103 and/or the server 105 .
  • terminal devices, networks and servers in Fig. 1 are only illustrative. According to the implementation needs, there can be any number of terminal devices, networks and servers.
  • Fig. 2 schematically shows a flowchart of a fusion model training method according to an embodiment of the present disclosure.
  • the method includes operations S210-S230.
  • the training source image and the training template image are input into the fusion model to obtain a training fusion image.
  • attribute alignment transformation is performed on the training fused image to obtain a training alignment image, wherein the attribute information of the training alignment image is consistent with the attribute information of the training source image.
  • the fusion model is trained using an identity loss function, wherein the identity loss function is generated for the training source image and the training alignment image.
  • the training source image may be an image to be fused, and the training source image may include a source human face object, but is not limited thereto, and may also include a source animal face object or other source objects.
  • the training template image may be a target image
  • the training template image may include a target human face object, but is not limited thereto, and may also include a target animal face object or other target objects.
  • the number of training template images is not limited. For example, there may be one or a plurality of them. As long as it can be input into the fusion model simultaneously with the training source image to obtain the training fusion image.
  • the training source image and the training template image may be fused by using the fusion model to generate a training fusion image.
  • the identity information of the training source image is transferred to the training template image, while keeping the attribute information of the training template image unchanged.
  • the fusion model can be trained by constraining the identity similarity between the identity information of the training alignment image and the identity information of the training source image. For example, generate an identity loss function for training aligned images and training source images, and use the identity loss function to train a fusion model.
  • attribute alignment transformation can be performed on the attribute information of the training fused image to generate the training alignment image by means of attribute alignment transformation.
  • the attribute information of the training aligned images is consistent with that of the training source images.
  • the identity loss function to calculate the identity loss value between the identity information of the training alignment image and the identity information of the training source image, the interference of the attribute information between the two has been eliminated, and only the identity information is involved. Therefore, in the process of using the identity loss function to train the fusion model, the anti-noise introduced due to the inconsistency of attribute information will not be generated, thereby improving the feasibility and stability of the fusion model training.
  • the training source images and training template images related to human face objects are obtained through various open, legal and compliant ways, for example, they can come from public data sets, or through the user’s licensed images.
  • the fusion model is not a fusion model for a specific user, and cannot reflect the personal information of a specific user.
  • the construction of the fusion model is performed after authorization by the user, and its construction process complies with relevant laws and regulations.
  • attribute alignment transformation is performed on the training fused image to obtain the training alignment image.
  • the attribute alignment transformation may include, for example, one or more of posture attribute alignment transformation, makeup attribute alignment transformation, and expression attribute alignment transformation. But it is not limited to this. It can also include age attribute alignment transformation, head shape attribute alignment transformation, etc.
  • the pose attribute alignment transformation may be changing the face pose, for example, simulating different face poses, and transforming the face into a frontal face.
  • the makeup attribute alignment transformation may be to change the makeup, for example, the migration of the makeup.
  • expression attribute alignment transformation may refer to changing the expression of a face, including the expression of image regions meaningful for synthesizing expressions, such as lips and nose.
  • various attribute transformation networks can be used to perform attribute alignment transformation on the training fused images.
  • a multi-attribute aligned transformation model may be utilized.
  • a multi-attribute alignment transformation model formed by combining StyleGAN (style attribute generation confrontation network) with 3DMM (3D Morphable Model, three-dimensional variable face model).
  • Using the multi-attribute alignment transformation model provided by the embodiment of the present disclosure to perform attribute alignment transformation on the training fused image can rapidly perform editing processing of attribute alignment changes on various attribute information.
  • the generated training alignment image and the training source image can simultaneously satisfy the consistency of posture attribute information, makeup attribute information, and expression attribute information.
  • the attribute feature vectors of the training fusion image and the training source image can be simultaneously input as input data into the multi-attribute alignment transformation model to obtain the training alignment image processed by the attribute alignment transformation.
  • the training alignment image is constrained by the attribute feature vector of the training source image, so that the attribute information of the training alignment image is consistent with the attribute information of the training source image.
  • the identity loss value obtained by using the identity information of the training alignment image and the identity information of the training source image will not introduce additional attribute information, reduce the interference of attribute information, and improve the training success rate of fused images.
  • each training sample may include: a training source image and a training template image.
  • the source object in the training source image and the target object in the training template image may have the same category (for example, the same category as human face or animal face), different attribute information, and different identity information.
  • the identity loss function between the identity information of the training alignment image and the identity information of the training source image can be calculated using the identity loss function generated for the training alignment image and the training source image, and the fusion can be adjusted based on the identity loss value parameters of the model until the identity loss value satisfies a predetermined identity loss threshold.
  • a fusion model whose identity loss value satisfies a predetermined identity loss threshold is used as a trained fusion model, for example, a fusion model whose identity loss value is greater than or equal to a predetermined identity loss threshold is used as a trained fusion model, so that the trained fusion model is used as an image fusion model application model.
  • a fusion model may also be trained by using a combination of an identity loss function and an attribute loss function.
  • the identity information of the training alignment image can be consistent with the identity information of the training source image, and the attribute information of the training fusion image can be consistent with the attribute information of the training template image.
  • FIG. 3 schematically shows a flowchart of a method for training a fusion model according to another embodiment of the present disclosure.
  • a training source image 310 and a training template image 320 may be input into a fusion model 330 to obtain a training fusion image 340 .
  • Perform attribute alignment transformation on the training fusion image 340 for example, input the training fusion image 340 to the attribute transformation network 350 to obtain the training alignment image 360 .
  • Identity loss function 370 is generated for training source images 310 and training aligned images 360 .
  • An attribute loss function 380 is generated for the training fused image 340 and the training template image 320 . Based on the identity loss function 370 and the attribute loss function 380, a joint loss function is determined. The fusion model is trained using a joint loss function.
  • the attribute loss function may be a feature matching loss function (GAN Feature Matching) in the generative adversarial network series, but it is not limited thereto, and may also be other feature matching loss functions. As long as it is a loss function that can be used to constrain the attribute consistency between the attribute information of the training template image and the attribute information of the training fusion image.
  • GAN Feature Matching GAN Feature Matching
  • the identity loss function may be an ArcFace (arc surface) loss function, but it is not limited thereto, and may also be other feature matching loss functions. As long as it is a loss function that can be used to constrain the identity consistency between the identity information of the training source image and the identity information of the training alignment image.
  • ArcFace arc surface loss function
  • the joint loss function L may be determined by combining, eg, adding, the attribute loss function L1 and the identity loss function L2 .
  • L L 1 +L 2 .
  • using the joint loss function to train the fusion model may include the following operations.
  • training the fusion model may include the following operations.
  • a joint loss value is obtained.
  • the joint loss value is compared with a predetermined joint loss threshold, and when the joint loss value does not meet the predetermined joint loss threshold, the parameters of the fusion model can be adjusted.
  • the joint loss value satisfies the predetermined joint loss threshold, for example, when the joint loss value is greater than or equal to the predetermined joint loss threshold, it may indicate that the fusion model training is completed.
  • a joint loss value is obtained.
  • the parameters of the fusion model are adjusted until the joint loss value converges. In the case where the joint loss value converges, it indicates that the training of the fused model is complete.
  • the identity information of the training alignment image may be consistent with the identity information of the training source image
  • the attribute information of the training fusion image output by the trained fusion model may be consistent with the attribute information of the training template image, thereby making the training fusion
  • the image retains identity similarity with the fused training source image, and the training fused image retains attribute similarity with the training template image.
  • the fusion model can be used as a generator, combined with a discriminator, and the fusion model can be further trained by using a training method of a Generative Adversarial Network.
  • GAN Generative Adversarial Network
  • the discriminator can be constructed based on a neural network.
  • a neural network For example, deep neural network (Deep Neural Network, DNN), convolutional neural network (Convolutional Neural Network, CNN), recurrent neural network (Recurrent Neural Network, RNN), etc., are not limited here, as long as they can be matched with the fusion model to achieve generation against the network.
  • the training process of the generative adversarial network may include the following operations.
  • the parameters of the fusion model can be fixed and the discriminator trained.
  • the training fusion image and the training source image output by the fusion model can be used as the discrimination training data of the discriminator, and the discriminator can be trained with the discrimination training data.
  • the fusion model is trained once, so that the discriminator cannot distinguish the training fusion image from the training source image as much as possible.
  • the output probability of the discriminator is 0.5, which means that the fusion model training is completed.
  • the fusion model is trained as a generator by generating an adversarial network to improve the authenticity of the fusion image output by the fusion model, so that the fusion image fits the real image.
  • Fig. 4 schematically shows a flowchart of a method for training a fusion model according to another embodiment of the present disclosure.
  • the difference between the training method of the fusion model shown in FIG. 4 and the training method of the fusion model shown in FIG. 3 is that the key point alignment preprocessing is performed on the training source image and the training template image respectively.
  • the training source image 411 can be keypoint aligned to obtain the training alignment source image 412
  • the training template image 421 can be keypoint aligned to obtain the training alignment template image 422
  • the training alignment source image 412 and the training alignment template image 422 input into the fusion model 330 to obtain a training fusion image 340 .
  • Perform attribute alignment transformation on the training fusion image 340 for example, input the training fusion image 340 to the attribute transformation network 350 to obtain the training alignment image 360 .
  • Identity loss function 370 is generated for training aligned source image 412 and training aligned image 360 .
  • Attribute loss function 380 is generated for training fused image 340 and training aligned template image 422 . Based on the identity loss function 370 and the attribute loss function 380, a joint loss function is determined. The fusion model is trained using a joint loss function.
  • the training source image may be detected using 5 face key points, and then the cropped key point-aligned training alignment source image may be obtained through ArcFace cropping.
  • the training template image can be detected using 72 face key points, and then the cropped key point alignment can be obtained by FFHQ (Flickr-Faces-High-Quality, high-definition face data set) cropping method
  • FFHQ flickr-Faces-High-Quality, high-definition face data set
  • the image key point alignment of the two images input to the fusion model is consistent, which is conducive to the generation of the fusion model training fusion image and speeds up the fusion model. training speed.
  • Fig. 5 schematically shows a flowchart of an image fusion method according to an embodiment of the present disclosure.
  • the method includes operation S510.
  • the image to be fused and the template image are input into the fusion model to obtain a fusion image, wherein the fusion model is trained by using the fusion model training method provided by the embodiment of the present disclosure.
  • the image to be fused may include a source human face object, but is not limited thereto, and may also include a source animal face object or other source objects.
  • the template image may be a target image, and the template image may include a target human face object, but is not limited thereto, and may also include a target animal face object, or other target objects.
  • the number of template images is not limited. For example, there may be one or a plurality of them. As long as it can be input into the fusion model simultaneously with the image to be fused to obtain the fused image.
  • the image to be fused and the template image can be fused to generate a fused image.
  • Using the image fusion method provided by the embodiment of the present disclosure using the fusion model trained by the fusion model training method provided by the embodiment of the present disclosure to generate a fusion image, improve the identity similarity between the fusion image and the image to be fusion, and reduce the number of fusion images Artifacts and other issues caused by the interference of attribute information.
  • FIG. 6 the image fusion method shown in FIG. 5 will be further described in conjunction with specific embodiments.
  • Fig. 6 schematically shows a flowchart of an image fusion method according to another embodiment of the present disclosure.
  • the image 611 to be fused is keypoint aligned to obtain an image 612 to be fused;
  • the template image 621 is keypoint aligned to obtain an alignment template image 622; and the alignment image 612 to be fused and the alignment template image 622 are input Into the fusion model 630, a fusion image 640 is obtained.
  • the image to be fused can be detected by using 5 key points of the face, and then the cropped image to be fused that is aligned with the key points can be obtained through ArcFace cropping.
  • the template image can be detected using 72 key points of the face, and then obtained by FFHQ (Flickr-Faces-High-Quality, high-definition face data set) cropping method to obtain the aligned key points Align the template image.
  • FFHQ lickr-Faces-High-Quality, high-definition face data set
  • the image key point alignment of the two images input to the fusion model is consistent, which is conducive to the generation of the fusion image of the fusion model, speeds up the processing speed, and Improve the realism of fused images.
  • Fig. 7 schematically shows a block diagram of a fusion model training device according to an embodiment of the present disclosure.
  • an apparatus 700 for training a fusion model may include a training fusion module 710 , an attribute transformation module 720 , and a training module 730 .
  • the training fusion module 710 is configured to input the training source image and the training template image into the fusion model to obtain the training fusion image.
  • the attribute transformation module 720 is configured to perform attribute alignment transformation on the training fused image to obtain a training alignment image, wherein the attribute information of the training alignment image is consistent with the attribute information of the training source image.
  • the training module 730 is configured to train the fusion model using an identity loss function, wherein the identity loss function is generated for the training source image and the training alignment image.
  • the training module may include a joint unit and a training unit.
  • the joint unit is configured to determine the joint loss function based on the identity loss function and the attribute loss function, wherein the attribute loss function is generated for the training fusion image and the training template image.
  • a training unit is used to train the fusion model using a joint loss function.
  • the training unit may include a first acquisition subunit, a first input subunit, a second acquisition subunit, a second input subunit, and a training subunit.
  • the first obtaining subunit is used to obtain the first identity information of the training source image and the second identity information of the training alignment image.
  • the first input subunit is configured to input the first identity information and the second identity information into the identity loss function to obtain an identity loss value.
  • the second obtaining subunit is used to obtain the first attribute information of the training template image and the second attribute information of the training fusion image.
  • the second input subunit is configured to input the first attribute information and the second attribute information into the identity loss function to obtain an attribute loss value.
  • the training subunit is used to train the fusion model based on the identity loss value and the attribute loss value.
  • the training fusion module may include a first training alignment unit, a second training alignment unit, and a training fusion unit.
  • the first training alignment unit is configured to perform key point alignment on the training source image to obtain the training alignment source image.
  • the second training alignment unit is configured to perform key point alignment on the training template image to obtain the training alignment template image.
  • the training fusion unit is used to input the training alignment source image and the training alignment template image into the fusion model to obtain the training fusion image.
  • the attribute alignment transformation includes at least one of the following: gesture attribute alignment transformation, makeup attribute alignment transformation, and expression attribute alignment transformation.
  • Fig. 8 schematically shows a block diagram of an image fusion device according to an embodiment of the present disclosure.
  • the image fusion apparatus 800 may include a fusion module 810 .
  • the fusion module 810 is configured to input the image to be fused and the template image into the fusion model to obtain a fused image.
  • the fusion model can be trained by using a fusion model training method.
  • the fusion module may include a first alignment unit, a second alignment unit, and a fusion unit.
  • the first alignment unit is configured to perform key point alignment on the images to be fused to obtain aligned images to be fused.
  • the second alignment unit is configured to perform key point alignment on the template image to obtain an aligned template image.
  • the fusion unit is configured to input the aligned image to be fused and the aligned template image into the fusion model to obtain a fused image.
  • the present disclosure also provides an electronic device, a readable storage medium, and a computer program product.
  • an electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein, the memory stores instructions executable by at least one processor, and the instructions are processed by at least one The processor is executed, so that at least one processor can perform the method as described above.
  • a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the method as described above.
  • a computer program product includes a computer program, and the computer program implements the above method when executed by a processor.
  • FIG. 9 shows a schematic block diagram of an example electronic device 900 that may be used to implement embodiments of the present disclosure.
  • Electronic device is intended to represent various forms of digital computers, such as laptops, desktops, workstations, personal digital assistants, servers, blade servers, mainframes, and other suitable computers.
  • Electronic devices may also represent various forms of mobile devices, such as personal digital processing, cellular telephones, smart phones, wearable devices, and other similar computing devices.
  • the components shown herein, their connections and relationships, and their functions, are by way of example only, and are not intended to limit implementations of the disclosure described and/or claimed herein.
  • the device 900 includes a computing unit 901 that can execute according to a computer program stored in a read-only memory (ROM) 902 or loaded from a storage unit 908 into a random-access memory (RAM) 903. Various appropriate actions and treatments. In the RAM 903, various programs and data necessary for the operation of the device 900 can also be stored.
  • the computing unit 901, ROM 902, and RAM 903 are connected to each other through a bus 904.
  • An input/output (I/O) interface 905 is also connected to the bus 904 .
  • the I/O interface 905 includes: an input unit 906, such as a keyboard, a mouse, etc.; an output unit 907, such as various types of displays, speakers, etc.; a storage unit 908, such as a magnetic disk, an optical disk, etc. ; and a communication unit 909, such as a network card, a modem, a wireless communication transceiver, and the like.
  • the communication unit 909 allows the device 900 to exchange information/data with other devices over a computer network such as the Internet and/or various telecommunication networks.
  • the computing unit 901 may be various general-purpose and/or special-purpose processing components having processing and computing capabilities. Some examples of computing units 901 include, but are not limited to, central processing units (CPUs), graphics processing units (GPUs), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, digital signal processing processor (DSP), and any suitable processor, controller, microcontroller, etc.
  • the computing unit 901 executes various methods and processes described above, such as an image fusion method or a fusion model training method.
  • the image fusion method or the training method of the fusion model may be implemented as a computer software program tangibly embodied in a machine-readable medium, such as the storage unit 908 .
  • part or all of the computer program may be loaded and/or installed on the device 900 via the ROM 902 and/or the communication unit 909.
  • the computer program When the computer program is loaded into the RAM 903 and executed by the computing unit 901, one or more steps of the image fusion method or the training method of the fusion model described above can be performed.
  • the computing unit 901 may be configured in any other appropriate manner (for example, by means of firmware) to execute an image fusion method or a fusion model training method.
  • Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips Implemented in a system of systems (SOC), load programmable logic device (CPLD), computer hardware, firmware, software, and/or combinations thereof.
  • FPGAs field programmable gate arrays
  • ASICs application specific integrated circuits
  • ASSPs application specific standard products
  • SOC system of systems
  • CPLD load programmable logic device
  • computer hardware firmware, software, and/or combinations thereof.
  • programmable processor can be special-purpose or general-purpose programmable processor, can receive data and instruction from storage system, at least one input device, and at least one output device, and transmit data and instruction to this storage system, this at least one input device, and this at least one output device an output device.
  • Program codes for implementing the methods of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special purpose computer, or other programmable data processing devices, so that the program codes, when executed by the processor or controller, make the functions/functions specified in the flow diagrams and/or block diagrams Action is implemented.
  • the program code may execute entirely on the machine, partly on the machine, as a stand-alone software package partly on the machine and partly on a remote machine or entirely on the remote machine or server.
  • a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device.
  • a machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium.
  • a machine-readable medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing.
  • machine-readable storage media would include one or more wire-based electrical connections, portable computer discs, hard drives, random access memory (RAM), read only memory (ROM), erasable programmable read only memory (EPROM or flash memory), optical fiber, compact disk read only memory (CD-ROM), optical storage, magnetic storage, or any suitable combination of the foregoing.
  • RAM random access memory
  • ROM read only memory
  • EPROM or flash memory erasable programmable read only memory
  • CD-ROM compact disk read only memory
  • magnetic storage or any suitable combination of the foregoing.
  • the systems and techniques described herein can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user. ); and a keyboard and pointing device (eg, a mouse or a trackball) through which a user can provide input to the computer.
  • a display device e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor
  • a keyboard and pointing device eg, a mouse or a trackball
  • Other kinds of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and can be in any form (including Acoustic input, speech input or, tactile input) to receive input from the user.
  • the systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., as a a user computer having a graphical user interface or web browser through which a user can interact with embodiments of the systems and techniques described herein), or including such backend components, middleware components, Or any combination of front-end components in a computing system.
  • the components of the system can be interconnected by any form or medium of digital data communication, eg, a communication network. Examples of communication networks include: Local Area Network (LAN), Wide Area Network (WAN) and the Internet.
  • a computer system may include clients and servers.
  • Clients and servers are generally remote from each other and typically interact through a communication network.
  • the relationship of client and server arises by computer programs running on the respective computers and having a client-server relationship to each other.
  • the server can be a cloud server, a server of a distributed system, or a server combined with a blockchain.
  • steps may be reordered, added or deleted using the various forms of flow shown above.
  • each step described in the present disclosure may be executed in parallel, sequentially, or in a different order, as long as the desired result of the technical solution disclosed in the present disclosure can be achieved, no limitation is imposed herein.

Landscapes

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

Abstract

本公开提供了融合模型的训练方法、图像融合方法、装置、电子设备、存储介质以及程序产品，涉及人工智能技术领域，尤其涉及计算机视觉和深度学习技术领域，可应用于人脸图像处理和人脸识别等场景。具体实现方案为：将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像；对训练融合图像进行属性对齐变换，得到训练对齐图像，训练对齐图像的属性信息和训练源图像的属性信息一致；以及利用身份损失函数训练融合模型，身份损失函数是针对训练源图像和训练对齐图像生成的。

Description

融合模型的训练方法、图像融合方法、装置、设备及介质
本申请要求于2021年09月30日递交的中国专利申请No.202111168236.2的优先权，其内容一并在此作为参考。
技术领域
本公开涉及人工智能技术领域，尤其涉及计算机视觉和深度学习技术领域，可应用于人脸图像处理和人脸识别等场景，具体涉及融合模型的训练方法、图像融合方法、装置、电子设备、存储介质以及程序产品。
背景技术
图像融合可以是指将两幅或者多幅图像综合成一幅新的图像的技术。图像融合能够利用多幅图像之间的相关性和互补性，使得融合后得到的新的图像有更全面、清晰的内容展示，从而有利于识别和探测。为公共安全、信息安全、金融安全层面的应用发展提供巨大帮助。
发明内容
本公开提供了一种融合模型的训练方法、图像融合方法、装置、电子设备、存储介质以及程序产品。
根据本公开的一方面，提供了一种融合模型的训练方法，包括：将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像；对所述训练融合图像进行属性对齐变换，得到训练对齐图像，其中，所述训练对齐图像的属性信息和所述训练源图像的属性信息一致；以及利用身份损失函数训练所述融合模型，其中，所述身份损失函数是针对所述训练源图像和所述训练对齐图像生成的。
根据本公开的另一方面，提供了一种图像融合方法，包括：将待融合图像和模板图像输入至融合模型中，得到融合图像；其中，所述融合模型利用根据如上所述的融合模型的训练方法训练得到。
根据本公开的另一方面，提供了一种融合模型的训练装置，包括：训练融合模块，用于将训练源图像和训练模板图像输入至融合模型中，得到 训练融合图像；属性变换模块，用于对所述训练融合图像进行属性对齐变换，得到训练对齐图像，其中，所述训练对齐图像的属性信息和所述训练源图像的属性信息一致；以及训练模块，用于利用身份损失函数训练所述融合模型，其中，所述身份损失函数是针对所述训练源图像和所述训练对齐图像生成的。
根据本公开的另一方面，提供了一种图像融合装置，包括：融合模块，用于将待融合图像和模板图像输入至融合模型中，得到融合图像；其中，所述融合模型利用如上所述的融合模型的训练方法训练得到。
根据本公开的另一方面，提供了一种电子设备，包括：至少一个处理器；以及与所述至少一个处理器通信连接的存储器；其中，所述存储器存储有可被所述至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行如上所述的方法。
根据本公开的另一方面，提供了一种存储有计算机指令的非瞬时计算机可读存储介质，其中，所述计算机指令用于使所述计算机执行如上所述的方法。
根据本公开的另一方面，提供了一种计算机程序产品，包括计算机程序，所述计算机程序在被处理器执行时实现如上所述的方法。
应当理解，本部分所描述的内容并非旨在标识本公开的实施例的关键或重要特征，也不用于限制本公开的范围。本公开的其它特征将通过以下的说明书而变得容易理解。
附图说明
附图用于更好地理解本方案，不构成对本公开的限定。其中：
图1示意性示出了根据本公开实施例的可以应用图像融合方法及装置的示例性系统架构；
图2示意性示出了根据本公开实施例的融合模型的训练方法的流程图；
图3示意性示出了根据本公开另一实施例的融合模型的训练方法的流程示意图；
图4示意性示出了根据本公开另一实施例的融合模型的训练方法的流程示意图；
图5示意性示出了根据本公开实施例的图像融合方法的流程图；
图6示意性示出了根据本公开另一实施例的图像融合方法的流程图；
图7示意性示出了根据本公开实施例的融合模型的训练装置的框图；
图8示意性示出了根据本公开实施例的图像融合装置的框图；以及
图9示意性示出了根据本公开实施例的适于实现图像融合方法的电子设备的框图。
具体实施方式
以下结合附图对本公开的示范性实施例做出说明，其中包括本公开实施例的各种细节以助于理解，应当将它们认为仅仅是示范性的。因此，本领域普通技术人员应当认识到，可以对这里描述的实施例做出各种改变和修改，而不会背离本公开的范围和精神。同样，为了清楚和简明，以下的描述中省略了对公知功能和结构的描述。
本公开提供了一种融合模型的训练方法、图像融合方法、装置、电子设备、存储介质以及程序产品。
根据本公开的实施例，融合模型的训练方法可以包括：将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像；对训练融合图像进行属性对齐变换，得到训练对齐图像，其中，训练对齐图像的属性信息和训练源图像的属性信息一致；以及利用身份损失函数训练融合模型，其中，身份损失函数是针对训练源图像和训练对齐图像生成的。
根据本公开的实施例，图像融合的方法，将待融合图像和模板图像输入至融合模型中，得到融合图像，其中，融合模型利用本公开实施例提供的融合模型的训练方法训练得到。
本公开的技术方案中，所涉及的用户个人信息的收集、存储、使用、加工、传输、提供、公开和应用等处理，均符合相关法律法规的规定，采取了必要保密措施，且不违背公序良俗。
在本公开的技术方案中，在获取或采集用户个人信息之前，均获取了用户的授权或同意。
图1示意性示出了根据本公开实施例的可以应用图像融合方法及装置的示例性系统架构。
需要注意的是，图1所示仅为可以应用本公开实施例的系统架构的示例，以帮助本领域技术人员理解本公开的技术内容，但并不意味着本公开实施例不可以用于其他设备、系统、环境或场景。例如，在另一实施例中，可以应用图像融合方法及装置的示例性系统架构可以包括终端设备，但终端设备可以无需与服务器进行交互，即可实现本公开实施例提供的图像融合方法及装置。
如图1所示，根据该实施例的系统架构100可以包括终端设备101、102、103，网络104和服务器105。网络104用以在终端设备101、102、103和服务器105之间提供通信链路的介质。网络104可以包括各种连接类型，例如有线和/或无线通信链路等等。
用户可以使用终端设备101、102、103通过网络104与服务器105交互，以发送源图像和模板图像，接收融合图像。终端设备101、102、103上可以安装有各种通讯客户端应用，例如加载有图像融合方法的应用程序等(仅为示例)。
终端设备101、102、103可以是具有显示屏并且具有摄像装置的各种电子设备，包括但不限于智能手机、平板电脑、膝上型便携计算机和台式计算机等等。
服务器105可以是提供各种服务的服务器，例如对用户利用终端设备101、102、103所上传的源图像和模板图像提供支持的后台管理服务器(仅为示例)。后台管理服务器可以对源图像和模板图像进行图像融合处理，得到融合图像，并将融合图像反馈给终端设备。
需要说明的是，本公开实施例所提供的图像融合方法一般可以由终端设备101、102、或103执行。相应地，本公开实施例所提供的图像融合装置也可以设置于终端设备101、102、或103中。
或者，本公开实施例所提供的图像融合方法一般也可以由服务器105执行。相应地，本公开实施例所提供的图像融合装置一般可以设置于服务器105中。本公开实施例所提供的图像融合方法也可以由不同于服务器105且能够与终端设备101、102、103和/或服务器105通信的服务器或服务器集群执行。相应地，本公开实施例所提供的图像融合装置也可以设置于不同于服务器105且能够与终端设备101、102、103和/或服务器105 通信的服务器或服务器集群中。
应该理解，图1中的终端设备、网络和服务器的数目仅仅是示意性的。根据实现需要，可以具有任意数目的终端设备、网络和服务器。
图2示意性示出了根据本公开实施例的融合模型的训练方法的流程图。
如图2所示，该方法包括操作S210～S230。
在操作S210，将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像。
在操作S220，对训练融合图像进行属性对齐变换，得到训练对齐图像，其中，训练对齐图像的属性信息和训练源图像的属性信息一致。
在操作S230，利用身份损失函数训练融合模型，其中，身份损失函数是针对训练源图像和训练对齐图像生成的。
根据本公开的实施例，训练源图像可以是待融合的图像，训练源图像中可以包括源人脸对象，但是并不局限于此，还可以包括源动物脸对象、或者其他源对象。
根据本公开的实施例，训练模板图像可以是目标图像，训练模板图像中可以包括目标人脸对象，但是并不局限于此，还可以包括目标动物脸对象、或者其他目标对象。
需要说明的是，训练模板图像的数量不做限定。例如可以是1个，也可以是多个。只要是能够与训练源图像同时输入至融合模型中，得到训练融合图像即可。
根据本公开的实施例，利用融合模型，可以将训练源图像和训练模板图像进行融合，生成训练融合图像。
例如，利用融合模型，将训练源图像的身份信息迁移到训练模板图像中，同时保留训练模板图像的属性信息不变。
根据本公开的实施例，可以通过约束训练对齐图像的身份信息与训练源图像的身份信息之间的身份相似度来训练融合模型。例如，生成针对训练对齐图像与训练源图像的身份损失函数，利用身份损失函数来训练融合模型。
根据本公开的实施例，可以利用属性对齐变换的方式，对训练融合图 像的属性信息进行属性对齐变换，生成训练对齐图像。训练对齐图像的属性信息与训练源图像的属性信息一致。进而在利用身份损失函数来计算训练对齐图像的身份信息与训练源图像的身份信息之间的身份损失值的情况下，已经将二者之间的属性信息的干扰排出，仅涉及身份信息。由此可以使得在利用身份损失函数训练融合模型的过程中，不会产生因属性信息不一致而引入的对抗噪声，进而提高融合模型训练的可实施性以及稳定性。
在本公开的实施例中，涉及人脸对象的训练源图像和训练模板图像通过各种公开、合法合规的方式获取，例如可以来自于公开数据集，或者是经过了人脸图像对应的用户的授权的图像。
需要说明的是，在本公开的实施例中融合模型不是针对某一特定用户的融合模型，并不能反映出某一特定用户的个人信息。融合模型的构建是在经用户授权后执行的，其构建过程符合相关法律法规。
下面结合具体实施例，并参考图3～图6对例如图2所示的融合模型的训练方法做进一步说明。
根据本公开的实施例，操作S220中，对训练融合图像进行属性对齐变换，得到训练对齐图像。其中，属性对齐变换可以包括例如姿态属性对齐变换、妆容属性对齐变换、表情属性对齐变换中的一项或多项。但是并不局限于此。还可以包括年龄属性对齐变换、头型属性对齐变换等。
根据本公开的实施例，姿态属性对齐变换可以是更改脸部姿态，例如仿真不同的脸部姿态，将脸部进行正脸化等变换。
根据本公开的实施例，妆容属性对齐变换可以是更改妆容，例如妆容的迁移。
根据本公开的实施例，表情属性对齐变换可以指更改脸部的表情，包括嘴唇、鼻子等对合成表情有意义的图像区域的表情。
根据本公开的实施例，可以采用各种属性变换网络对训练融合图像进行属性对齐变换。
根据本公开的示例性实施例，可以利用多属性的对齐变换模型。例如，利用StyleGAN(风格属性生成对抗网络)与3DMM(3D Morphable Model，三维可变人脸模型)结合而形成的多属性的对齐变换模型。
利用本公开实施例提供的多属性的对齐变换模型来对训练融合图像进行属性对齐变换，能够快速地对多种属性信息进行属性对齐变化的编辑处理。使得生成的训练对齐图像与训练源图像能够同时满足姿态属性信息一致、妆容属性信息一致、以及表情属性信息一致。
根据本公开的实施例，可以同时将训练融合图像和训练源图像的属性特征向量作为输入数据输入至多属性的对齐变换模型中，得到经属性对齐变换处理的训练对齐图像。训练对齐图像经训练源图像的属性特征向量约束，使得训练对齐图像的属性信息与训练源图像的属性信息保持一致。进而利用训练对齐图像的身份信息与训练源图像的身份信息得到的身份损失值不会引入额外的属性信息，降低属性信息的干扰，提高融合图像的训练成功率。
根据本公开的实施例，操作S230中，在利用身份损失函数训练融合模型的过程中，可以根据实际需要获取多个训练样本，每个训练样本可以包括：训练源图像和训练模板图像。其中，训练源图像中的源对象和训练模板图像中的目标对象之间可以类别相同(例如同为人脸或者动物脸的类别)、属性信息不同、以及身份信息不同。
根据本公开的实施例，可以利用针对训练对齐图像与训练源图像生成的身份损失函数来计算训练对齐图像的身份信息与训练源图像的身份信息之间的身份损失值，基于身份损失值调整融合模型的参数，直至身份损失值满足预定身份损失阈值。将身份损失值满足预定身份损失阈值的融合模型作为经训练的融合模型，例如身份损失值大于或等于预定身份损失阈值的融合模型作为经训练的融合模型，以便将经训练的融合模型作为图像融合的应用模型。
根据本公开的示例性实施例，还可以利用身份损失函数和属性损失函数结合的方式来训练融合模型。使得训练对齐图像的身份信息能够与训练源图像的身份信息保持一致，并且训练融合图像的属性信息能够与训练模板图像的属性信息保持一致。
图3示意性示出了根据本公开另一实施例的融合模型的训练方法的流程示意图。
如图3所示，可以将训练源图像310和训练模板图像320输入至融合模型330中，得到训练融合图像340。对训练融合图像340进行属性对齐变换，例如可以是将训练融合图像340输入至属性变换网络350得到训练对齐图像360。针对训练源图像310和训练对齐图像360生成身份损失函数370。针对训练融合图像340和训练模板图像320生成属性损失函数380。基于身份损失函数370和属性损失函数380，确定联合损失函数。利用联合损失函数训练融合模型。
根据本公开的实施例，属性损失函数可以是生成对抗网络系列中的特征匹配损失函数(GAN Feature Matching)，但是并不局限于此，还可以是其他特征匹配损失函数。只要是能够用来约束训练模板图像的属性信息与训练融合图像属性信息之间的属性一致性的损失函数即可。
根据本公开的实施例，身份损失函数可以是ArcFace(弧面)损失函数，但是并不局限于此，还可以是其他特征匹配损失函数。只要是能够用来约束训练源图像的身份信息与训练对齐图像身份信息之间的身份一致性的损失函数即可。
根据本公开的实施例，联合损失函数L可以是属性损失函数L 1与身份损失函数L 2结合例如相加确定。例如，L＝L 1+L 2。但是并不局限于此。还可以为属性损失函数与身份损失函数配置权重，将属性损失函数与身份损失函数并结合各自对应的权重W 1和W 2来确定联合损失函数。例如，L＝W 1*L 1+W 2*L 2。
根据本公开的实施例，利用联合损失函数训练融合模型可以包括如下操作。
例如，获取训练源图像的第一身份信息和训练对齐图像的第二身份信息；将第一身份信息和第二身份信息输入至身份损失函数中，得到身份损失值；获取训练模板图像的第一属性信息和训练融合图像的第二属性信息；将第一属性信息和第二属性信息输入至身份损失函数中，得到属性损失值；以及基于身份损失值和属性损失值，训练融合模型。
根据本公开的实施例，基于身份损失值和属性损失值，训练融合模型可以包括如下操作。
例如，基于身份损失值与属性损失值，得到联合损失值。将联合损失值与预定联合损失阈值进行比较，在联合损失值不满足预定联合损失阈值的情况下，可以调整融合模型的参数。在联合损失值满足预定联合损失阈值的情况下，例如在联合损失值大于或等于预定联合损失阈值的情况下，可以表明融合模型训练完成。
还例如，基于身份损失值与属性损失值，得到联合损失值。基于联合损失值，调整融合模型的参数，直至联合损失值收敛。在联合损失值收敛的情况下，表明融合模型的训练完成。
根据本公开的实施例，训练对齐图像的身份信息可以与训练源图像的身份信息一致，训练完成的融合模型输出的训练融合图像的属性信息可以与训练模板图像的属性信息一致，进而使得训练融合图像与融合训练源图像保留有身份相似性，并且训练融合图像与训练模板图像保留有属性相似性。
根据本公开的示例性实施例，可以基于生成对抗网络(GAN)，将融合模型作为生成器，再结合鉴别器，采用生成对抗网络的训练方式对融合模型做进一步地训练。
根据本公开的实施例，鉴别器可以基于神经网络构建。例如深度神经网络(Deep Neural Network，DNN)、卷积神经网络(Convolutional Neural Network，CNN)、循环神经网络(Recurrent NeuralNetwork，RNN)等，在此不做限定，只要能与融合模型匹配，实现生成对抗网络即可。
根据本公开的实施例，生成对抗网络的训练过程可以包括如下操作。例如，可以固定融合模型的参数，训练鉴别器。可以利用融合模型输出的训练融合图像和训练源图像作为鉴别器的鉴别训练数据，用鉴别训练数据来训练鉴别器。循环训练多次鉴别器后，训练一次融合模型，使得鉴别器尽可能区分不开训练融合图像和训练源图像。
根据本公开的实施例，经过多次训练迭代后，使得鉴别器的输出概率为0.5，即可认为融合模型训练完成。
根据本公开的实施例，融合模型作为生成器经过生成对抗网络的方式训练，提高融合模型输出的融合图像的真实度，使得融合图像贴合真实图 像。
图4示意性示出了根据本公开另一实施例的融合模型的训练方法的流程示意图。
如图4所示的融合模型的训练方法与如图3所示的融合模型的训练方法的不同之处在于，对训练源图像和训练模板图像分别进行关键点对齐预处理。例如，可以对训练源图像411进行关键点对齐，得到训练对齐源图像412；对训练模板图像421进行关键点对齐，得到训练对齐模板图像422；以及将训练对齐源图像412和训练对齐模板图像422输入至融合模型330中，得到训练融合图像340。对训练融合图像340进行属性对齐变换，例如可以是将训练融合图像340输入至属性变换网络350得到训练对齐图像360。针对训练对齐源图像412和训练对齐图像360生成身份损失函数370。针对训练融合图像340和训练对齐模板图像422生成属性损失函数380。基于身份损失函数370和属性损失函数380，确定联合损失函数。利用联合损失函数训练融合模型。
根据本公开的示例性实施例，可以对训练源图像利用脸部5点关键点检测，再通过ArcFace裁剪方式得到被裁剪的关键点对齐的训练对齐源图像。
根据本公开的示例性实施例，可以对训练模板图像利用脸部72点关键点检测，再通过FFHQ(Flickr-Faces-High-Quality，高清人脸数据集)裁剪方式得到被裁剪的关键点对齐的训练对齐模板图像。
利用本公开实施例提供的关键点对齐预处理操作，使得输入至融合模型的两个图像的图像关键点对齐例如分辨率等信息一致，有利于融合模型的训练融合图像的生成，加快融合模型的训练速度。而且，也有利于从训练对齐模板图像中提取属性信息，以及从训练对齐源图像中提取身份信息，方便身份损失值以及属性损失值的计算。
图5示意性示出了根据本公开实施例的图像融合方法的流程图。
如图5所示，该方法包括操作S510。
操作S510，将待融合图像和模板图像输入至融合模型中，得到融合图像，其中，融合模型利用本公开实施例提供的融合模型的训练方法训练 得到。
根据本公开的实施例，待融合图像中可以包括源人脸对象，但是并不局限于此，还可以包括源动物脸对象、或者其他源对象。
根据本公开的实施例，模板图像可以是目标图像，模板图像中可以包括目标人脸对象，但是并不局限于此，还可以包括目标动物脸对象、或者其他目标对象。
需要说明的是，模板图像的数量不做限定。例如可以是1个，也可以是多个。只要是能够与待融合图像同时输入至融合模型中，得到融合图像即可。
根据本公开的实施例，利用融合模型，可以将待融合图像和模板图像进行融合，生成融合图像。
利用本公开实施例提供的图像融合方法，利用本公开实施例提供的融合模型的训练方法训练得到的融合模型来生成融合图像，提高融合图像与待融合图像的身份相似度，并且可以减少融合图像中由于属性信息干扰而导致的伪影等问题。
下面参考图6，结合具体实施例对例如图5所示的图像融合方法做进一步说明。
图6示意性示出了根据本公开另一实施例的图像融合方法的流程示意图。
如图6所示，对待融合图像611进行关键点对齐，得到对齐待融合图像612；对模板图像621进行关键点对齐，得到对齐模板图像622；以及将对齐待融合图像612和对齐模板图像622输入至融合模型630中，得到融合图像640。
根据本公开的示例性实施例，可以对待融合图像利用脸部5点关键点检测，再通过ArcFace裁剪方式得到被裁剪的关键点对齐的对齐待融合图像。
根据本公开的示例性实施例，可以对模板图像利用脸部72点关键点检测，再通过FFHQ(Flickr-Faces-High-Quality，高清人脸数据集)裁剪方式得到被裁剪的关键点对齐的对齐模板图像。
利用本公开实施例提供的关键点对齐预处理操作，使得输入至融合模型的两个图像的图像关键点对齐例如分辨率等信息一致，有利于融合模型的融合图像的生成，加快处理速度，并提高融合图像的真实度。
图7示意性示出了根据本公开实施例的融合模型的训练装置的框图。
如图7所示，融合模型的训练装置700可以包括训练融合模块710、属性变换模块720、训练模块730。
训练融合模块710，用于将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像。
属性变换模块720，用于对训练融合图像进行属性对齐变换，得到训练对齐图像，其中，训练对齐图像的属性信息和训练源图像的属性信息一致。
训练模块730，用于利用身份损失函数训练融合模型，其中，身份损失函数是针对训练源图像和训练对齐图像生成的。
根据本公开的实施例，训练模块可以包括联合单元、训练单元。
联合单元，用于基于身份损失函数和属性损失函数，确定联合损失函数，其中，属性损失函数是针对训练融合图像和训练模板图像生成的。
训练单元，用于利用联合损失函数训练融合模型。
根据本公开的实施例，训练单元可以包括第一获取子单元、第一输入子单元、第二获取子单元、第二输入子单元、训练子单元。
第一获取子单元，用于获取训练源图像的第一身份信息和训练对齐图像的第二身份信息。
第一输入子单元，用于将第一身份信息和第二身份信息输入至身份损失函数中，得到身份损失值。
第二获取子单元，用于获取训练模板图像的第一属性信息和训练融合图像的第二属性信息。
第二输入子单元，用于将第一属性信息和第二属性信息输入至身份损失函数中，得到属性损失值。
训练子单元，用于基于身份损失值和属性损失值，训练融合模型。
根据本公开的实施例，训练融合模块可以包括第一训练对齐单元、第 二训练对齐单元、训练融合单元。
第一训练对齐单元，用于对训练源图像进行关键点对齐，得到训练对齐源图像。
第二训练对齐单元，用于对训练模板图像进行关键点对齐，得到训练对齐模板图像。
训练融合单元，用于将训练对齐源图像和训练对齐模板图像输入至融合模型中，得到训练融合图像。
根据本公开的实施例，属性对齐变换包括以下至少一项：姿态属性对齐变换、妆容属性对齐变换、表情属性对齐变换。
图8示意性示出了根据本公开实施例的图像融合装置的框图。
如图8所示，图像融合装置800可以包括融合模块810。
融合模块810，用于将待融合图像和模板图像输入至融合模型中，得到融合图像。
根据本公开的实施例，融合模型可以利用融合模型的训练方法训练得到。
根据本公开的实施例，融合模块可以包括第一对齐单元、第二对齐单元、融合单元。
第一对齐单元，用于对待融合图像进行关键点对齐，得到对齐待融合图像。
第二对齐单元，用于对模板图像进行关键点对齐，得到对齐模板图像。
融合单元，用于将对齐待融合图像和对齐模板图像输入至融合模型中，得到融合图像。
根据本公开的实施例，本公开还提供了一种电子设备、一种可读存储介质和一种计算机程序产品。
根据本公开的实施例，一种电子设备，包括：至少一个处理器；以及与至少一个处理器通信连接的存储器；其中，存储器存储有可被至少一个处理器执行的指令，指令被至少一个处理器执行，以使至少一个处理器能够执行如上所述的方法。
根据本公开的实施例，一种存储有计算机指令的非瞬时计算机可读存 储介质，其中，计算机指令用于使计算机执行如上所述的方法。
根据本公开的实施例，一种计算机程序产品，包括计算机程序，计算机程序在被处理器执行时实现如上所述的方法。
图9示出了可以用来实施本公开的实施例的示例电子设备900的示意性框图。电子设备旨在表示各种形式的数字计算机，诸如，膝上型计算机、台式计算机、工作台、个人数字助理、服务器、刀片式服务器、大型计算机、和其它适合的计算机。电子设备还可以表示各种形式的移动装置，诸如，个人数字处理、蜂窝电话、智能电话、可穿戴设备和其它类似的计算装置。本文所示的部件、它们的连接和关系、以及它们的功能仅仅作为示例，并且不意在限制本文中描述的和/或者要求的本公开的实现。
如图9所示，设备900包括计算单元901，其可以根据存储在只读存储器(ROM)902中的计算机程序或者从存储单元908加载到随机访问存储器(RAM)903中的计算机程序，来执行各种适当的动作和处理。在RAM 903中，还可存储设备900操作所需的各种程序和数据。计算单元901、ROM 902以及RAM 903通过总线904彼此相连。输入/输出(I/O)接口905也连接至总线904。
设备900中的多个部件连接至I/O接口905，包括：输入单元906，例如键盘、鼠标等；输出单元907，例如各种类型的显示器、扬声器等；存储单元908，例如磁盘、光盘等；以及通信单元909，例如网卡、调制解调器、无线通信收发机等。通信单元909允许设备900通过诸如因特网的计算机网络和/或各种电信网络与其他设备交换信息/数据。
计算单元901可以是各种具有处理和计算能力的通用和/或专用处理组件。计算单元901的一些示例包括但不限于中央处理单元(CPU)、图形处理单元(GPU)、各种专用的人工智能(AI)计算芯片、各种运行机器学习模型算法的计算单元、数字信号处理器(DSP)、以及任何适当的处理器、控制器、微控制器等。计算单元901执行上文所描述的各个方法和处理，例如图像融合方法或者融合模型的训练方法。例如，在一些实施例中，图像融合方法或者融合模型的训练方法可被实现为计算机软件程序，其被有形地包含于机器可读介质，例如存储单元908。在一些实施例中，计算机程序的部分或者全部可以经由ROM 902和/或通信单元909而被载 入和/或安装到设备900上。当计算机程序加载到RAM 903并由计算单元901执行时，可以执行上文描述的图像融合方法或者融合模型的训练方法的一个或多个步骤。备选地，在其他实施例中，计算单元901可以通过其他任何适当的方式(例如，借助于固件)而被配置为执行图像融合方法或者融合模型的训练方法。
本文中以上描述的系统和技术的各种实施方式可以在数字电子电路系统、集成电路系统、场可编程门阵列(FPGA)、专用集成电路(ASIC)、专用标准产品(ASSP)、芯片上系统的系统(SOC)、负载可编程逻辑设备(CPLD)、计算机硬件、固件、软件、和/或它们的组合中实现。这些各种实施方式可以包括：实施在一个或者多个计算机程序中，该一个或者多个计算机程序可在包括至少一个可编程处理器的可编程系统上执行和/或解释，该可编程处理器可以是专用或者通用可编程处理器，可以从存储系统、至少一个输入装置、和至少一个输出装置接收数据和指令，并且将数据和指令传输至该存储系统、该至少一个输入装置、和该至少一个输出装置。
用于实施本公开的方法的程序代码可以采用一个或多个编程语言的任何组合来编写。这些程序代码可以提供给通用计算机、专用计算机或其他可编程数据处理装置的处理器或控制器，使得程序代码当由处理器或控制器执行时使流程图和/或框图中所规定的功能/操作被实施。程序代码可以完全在机器上执行、部分地在机器上执行，作为独立软件包部分地在机器上执行且部分地在远程机器上执行或完全在远程机器或服务器上执行。
在本公开的上下文中，机器可读介质可以是有形的介质，其可以包含或存储以供指令执行系统、装置或设备使用或与指令执行系统、装置或设备结合地使用的程序。机器可读介质可以是机器可读信号介质或机器可读储存介质。机器可读介质可以包括但不限于电子的、磁性的、光学的、电磁的、红外的、或半导体系统、装置或设备，或者上述内容的任何合适组合。机器可读存储介质的更具体示例会包括基于一个或多个线的电气连接、便携式计算机盘、硬盘、随机存取存储器(RAM)、只读存储器(ROM)、可擦除可编程只读存储器(EPROM或快闪存储器)、光纤、便捷式紧凑盘只读存储器(CD-ROM)、光学储存设备、磁储存设备、或上述内容的 任何合适组合。
为了提供与用户的交互，可以在计算机上实施此处描述的系统和技术，该计算机具有：用于向用户显示信息的显示装置(例如，CRT(阴极射线管)或者LCD(液晶显示器)监视器)；以及键盘和指向装置(例如，鼠标或者轨迹球)，用户可以通过该键盘和该指向装置来将输入提供给计算机。其它种类的装置还可以用于提供与用户的交互；例如，提供给用户的反馈可以是任何形式的传感反馈(例如，视觉反馈、听觉反馈、或者触觉反馈)；并且可以用任何形式(包括声输入、语音输入或者、触觉输入)来接收来自用户的输入。
可以将此处描述的系统和技术实施在包括后台部件的计算系统(例如，作为数据服务器)、或者包括中间件部件的计算系统(例如，应用服务器)、或者包括前端部件的计算系统(例如，具有图形用户界面或者网络浏览器的用户计算机，用户可以通过该图形用户界面或者该网络浏览器来与此处描述的系统和技术的实施方式交互)、或者包括这种后台部件、中间件部件、或者前端部件的任何组合的计算系统中。可以通过任何形式或者介质的数字数据通信(例如，通信网络)来将系统的部件相互连接。通信网络的示例包括：局域网(LAN)、广域网(WAN)和互联网。
计算机系统可以包括客户端和服务器。客户端和服务器一般远离彼此并且通常通过通信网络进行交互。通过在相应的计算机上运行并且彼此具有客户端-服务器关系的计算机程序来产生客户端和服务器的关系。服务器可以是云服务器，也可以是分布式系统的服务器，或者是结合了区块链的服务器。
应该理解，可以使用上面所示的各种形式的流程，重新排序、增加或删除步骤。例如，本发公开中记载的各步骤可以并行地执行也可以顺序地执行也可以不同的次序执行，只要能够实现本公开公开的技术方案所期望的结果，本文在此不进行限制。
上述具体实施方式，并不构成对本公开保护范围的限制。本领域技术人员应该明白的是，根据设计要求和其他因素，可以进行各种修改、组合、子组合和替代。任何在本公开的精神和原则之内所作的修改、等同替换和改进等，均应包含在本公开保护范围之内。

Claims (17)

  1. 一种融合模型的训练方法，包括：
    将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像；
    对所述训练融合图像进行属性对齐变换，得到训练对齐图像，其中，所述训练对齐图像的属性信息和所述训练源图像的属性信息一致；以及
    利用身份损失函数训练所述融合模型，其中，所述身份损失函数是针对所述训练源图像和所述训练对齐图像生成的。
  2. 根据权利要求1所述的方法，其中，所述利用身份损失函数训练所述融合模型包括：
    基于所述身份损失函数和属性损失函数，确定联合损失函数，其中，所述属性损失函数是针对所述训练融合图像和所述训练模板图像生成的；以及
    利用所述联合损失函数训练所述融合模型。
  3. 根据权利要求2所述的方法，其中，所述利用所述联合损失函数训练所述融合模型包括：
    获取所述训练源图像的第一身份信息和所述训练对齐图像的第二身份信息；
    将所述第一身份信息和所述第二身份信息输入至所述身份损失函数中，得到身份损失值；
    获取所述训练模板图像的第一属性信息和所述训练融合图像的第二属性信息；
    将所述第一属性信息和所述第二属性信息输入至所述身份损失函数中，得到属性损失值；以及
    基于所述身份损失值和所述属性损失值，训练所述融合模型。
  4. 根据权利要求1所述的方法，其中，所述将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像包括：
    对所述训练源图像进行关键点对齐，得到训练对齐源图像；
    对所述训练模板图像进行关键点对齐，得到训练对齐模板图像；以及
    将所述训练对齐源图像和所述训练对齐模板图像输入至所述融合模型中，得到所述训练融合图像。
  5. 根据权利要求1所述的方法，其中，所述属性对齐变换包括以下至少一项：
    姿态属性对齐变换、妆容属性对齐变换、表情属性对齐变换。
  6. 一种图像融合方法，包括：
    将待融合图像和模板图像输入至融合模型中，得到融合图像；
    其中，所述融合模型利用根据权利要求1-5任一项所述的融合模型的训练方法训练得到。
  7. 根据权利要求6所述的方法，其中，所述将待融合图像和模板图像输入至融合模型中，得到融合图像包括：
    对所述待融合图像进行关键点对齐，得到对齐待融合图像；
    对所述模板图像进行关键点对齐，得到对齐模板图像；以及
    将所述对齐待融合图像和所述对齐模板图像输入至所述融合模型中，得到所述融合图像。
  8. 一种融合模型的训练装置，包括：
    训练融合模块，用于将训练源图像和训练模板图像输入至融合模型中，得到训练融合图像；
    属性变换模块，用于对所述训练融合图像进行属性对齐变换，得到训练对齐图像，其中，所述训练对齐图像的属性信息和所述训练源图像的属性信息一致；以及
    训练模块，用于利用身份损失函数训练所述融合模型，其中，所述身份损失函数是针对所述训练源图像和所述训练对齐图像生成的。
  9. 根据权利要求8所述的装置，其中，所述训练模块包括：
    联合单元，用于基于所述身份损失函数和属性损失函数，确定联合损失函数，其中，所述属性损失函数是针对所述训练融合图像和所述训练模板图像生成的；以及
    训练单元，用于利用所述联合损失函数训练所述融合模型。
  10. 根据权利要求9所述的装置，其中，所述训练单元包括：
    第一获取子单元，用于获取所述训练源图像的第一身份信息和所述训练对齐图像的第二身份信息；
    第一输入子单元，用于将所述第一身份信息和所述第二身份信息输入至所述身份损失函数中，得到身份损失值；
    第二获取子单元，用于获取所述训练模板图像的第一属性信息和所述训练融合图像的第二属性信息；
    第二输入子单元，用于将所述第一属性信息和所述第二属性信息输入至所述身份损失函数中，得到属性损失值；以及
    训练子单元，用于基于所述身份损失值和所述属性损失值，训练所述融合模型。
  11. 根据权利要求8所述的装置，其中，所述训练融合模块包括：
    第一训练对齐单元，用于对所述训练源图像进行关键点对齐，得到训练对齐源图像；
    第二训练对齐单元，用于对所述训练模板图像进行关键点对齐，得到训练对齐模板图像；以及
    训练融合单元，用于将所述训练对齐源图像和所述训练对齐模板图像输入至所述融合模型中，得到所述训练融合图像。
  12. 根据权利要求8所述的装置，其中，所述属性对齐变换包括以下至少一项：
    姿态属性对齐变换、妆容属性对齐变换、表情属性对齐变换。
  13. 一种图像融合装置，包括：
    融合模块，用于将待融合图像和模板图像输入至融合模型中，得到融合图像；
    其中，所述融合模型利用根据权利要求1-5任一项所述的融合模型的训练方法训练得到。
  14. 根据权利要求13所述的装置，其中，所述融合模块包括：
    第一对齐单元，用于对所述待融合图像进行关键点对齐，得 到对齐待融合图像；
    第二对齐单元，用于对所述模板图像进行关键点对齐，得到对齐模板图像；以及
    融合单元，用于将所述对齐待融合图像和所述对齐模板图像输入至所述融合模型中，得到所述融合图像。
  15. 一种电子设备，包括：
    至少一个处理器；以及
    与所述至少一个处理器通信连接的存储器；其中，
    所述存储器存储有可被所述至少一个处理器执行的指令，所述指令被所述至少一个处理器执行，以使所述至少一个处理器能够执行权利要求1-5中任一项所述的融合模型的训练方法或者权利要求6-7中任一项所述的图像融合方法。
  16. 一种存储有计算机指令的非瞬时计算机可读存储介质，其中，所述计算机指令用于使所述计算机执行根据权利要求1-5中任一项所述的融合模型的训练方法或者权利要求6-7中任一项所述的图像融合方法。
  17. 一种计算机程序产品，包括计算机程序，所述计算机程序在被处理器执行时实现根据权利要求1-5中任一项所述的融合模型的训练方法或者权利要求6-7中任一项所述的图像融合方法。
PCT/CN2022/097872 2021-09-30 2022-06-09 融合模型的训练方法、图像融合方法、装置、设备及介质 WO2023050868A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202111168236.2A CN113902956B (zh) 2021-09-30 2021-09-30 融合模型的训练方法、图像融合方法、装置、设备及介质
CN202111168236.2 2021-09-30

Publications (1)

Publication Number Publication Date
WO2023050868A1 true WO2023050868A1 (zh) 2023-04-06

Family

ID=79190440

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/097872 WO2023050868A1 (zh) 2021-09-30 2022-06-09 融合模型的训练方法、图像融合方法、装置、设备及介质

Country Status (2)

Country Link
CN (1) CN113902956B (zh)
WO (1) WO2023050868A1 (zh)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113902956B (zh) * 2021-09-30 2023-04-07 北京百度网讯科技有限公司 融合模型的训练方法、图像融合方法、装置、设备及介质
CN114926322B (zh) * 2022-05-12 2024-03-15 北京百度网讯科技有限公司 图像生成方法、装置、电子设备和存储介质
CN115345782A (zh) * 2022-08-12 2022-11-15 腾讯科技（深圳）有限公司 图像处理方法、装置、计算机、可读存储介质及程序产品

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109146830A (zh) * 2018-07-17 2019-01-04 北京旷视科技有限公司 用于生成训练数据的方法、装置、系统和存储介质
US20200265219A1 (en) * 2017-09-18 2020-08-20 Board Of Trustees Of Michigan State University Disentangled representation learning generative adversarial network for pose-invariant face recognition
CN111783647A (zh) * 2020-06-30 2020-10-16 北京百度网讯科技有限公司 人脸融合模型的训练方法、人脸融合方法、装置及设备
CN111860167A (zh) * 2020-06-18 2020-10-30 北京百度网讯科技有限公司 人脸融合模型获取及人脸融合方法、装置及存储介质
CN113361387A (zh) * 2021-06-03 2021-09-07 湖南快乐阳光互动娱乐传媒有限公司 人脸图像融合方法及装置、存储介质及电子设备
CN113902956A (zh) * 2021-09-30 2022-01-07 北京百度网讯科技有限公司 融合模型的训练方法、图像融合方法、装置、设备及介质

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111353546B (zh) * 2020-03-09 2022-12-23 腾讯科技（深圳）有限公司 图像处理模型的训练方法、装置、计算机设备和存储介质
CN112766160B (zh) * 2021-01-20 2023-07-28 西安电子科技大学 基于多级属性编码器和注意力机制的人脸替换方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20200265219A1 (en) * 2017-09-18 2020-08-20 Board Of Trustees Of Michigan State University Disentangled representation learning generative adversarial network for pose-invariant face recognition
CN109146830A (zh) * 2018-07-17 2019-01-04 北京旷视科技有限公司 用于生成训练数据的方法、装置、系统和存储介质
CN111860167A (zh) * 2020-06-18 2020-10-30 北京百度网讯科技有限公司 人脸融合模型获取及人脸融合方法、装置及存储介质
CN111783647A (zh) * 2020-06-30 2020-10-16 北京百度网讯科技有限公司 人脸融合模型的训练方法、人脸融合方法、装置及设备
CN113361387A (zh) * 2021-06-03 2021-09-07 湖南快乐阳光互动娱乐传媒有限公司 人脸图像融合方法及装置、存储介质及电子设备
CN113902956A (zh) * 2021-09-30 2022-01-07 北京百度网讯科技有限公司 融合模型的训练方法、图像融合方法、装置、设备及介质

Also Published As

Publication number Publication date
CN113902956B (zh) 2023-04-07
CN113902956A (zh) 2022-01-07

Similar Documents

Publication Publication Date Title
WO2023050868A1 (zh) 融合模型的训练方法、图像融合方法、装置、设备及介质
WO2020006961A1 (zh) 用于提取图像的方法和装置
US20220084163A1 (en) Target image generation method and apparatus, server, and storage medium
WO2020103700A1 (zh) 一种基于微表情的图像识别方法、装置以及相关设备
US20220261968A1 (en) Image optimization method and apparatus, computer storage medium, and electronic device
US20210406579A1 (en) Model training method, identification method, device, storage medium and program product
US20220222925A1 (en) Artificial intelligence-based image processing method and apparatus, device, and storage medium
WO2021083069A1 (zh) 用于训练换脸模型的方法和设备
US20230071661A1 (en) Method for training image editing model and method for editing image
US20220189189A1 (en) Method of training cycle generative networks model, and method of building character library
JP7401606B2 (ja) 仮想オブジェクトリップ駆動方法、モデル訓練方法、関連装置及び電子機器
WO2022227765A1 (zh) 生成图像修复模型的方法、设备、介质及程序产品
US20230114293A1 (en) Method for training a font generation model, method for establishing a font library, and device
US20230143452A1 (en) Method and apparatus for generating image, electronic device and storage medium
WO2023024653A1 (zh) 图像处理方法、图像处理装置、电子设备以及存储介质
US20220092835A1 (en) Method and apparatus for recognizing face-swap, device and computer readable storage medium
US20230047748A1 (en) Method of fusing image, and method of training image fusion model
CN115880506B (zh) 图像生成方法、模型的训练方法、装置及电子设备
CN113052962A (zh) 模型训练、信息输出方法，装置，设备以及存储介质
CN114926322B (zh) 图像生成方法、装置、电子设备和存储介质
WO2024040870A1 (zh) 文本图像生成、训练、文本图像处理方法以及电子设备
WO2023116744A1 (zh) 图像处理方法、装置、设备及介质
US20230115765A1 (en) Method and apparatus of transferring image, and method and apparatus of training image transfer model
CN116402914A (zh) 用于确定风格化图像生成模型的方法、装置及产品
US20230154077A1 (en) Training method for character generation model, character generation method, apparatus and storage medium

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 18020891

Country of ref document: US

121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22874266

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE