WO2023272414A1 - Procédé de traitement d'image et appareil de traitement d'image - Google Patents

Procédé de traitement d'image et appareil de traitement d'image Download PDF

Info

Publication number
WO2023272414A1
WO2023272414A1 PCT/CN2021/102613 CN2021102613W WO2023272414A1 WO 2023272414 A1 WO2023272414 A1 WO 2023272414A1 CN 2021102613 W CN2021102613 W CN 2021102613W WO 2023272414 A1 WO2023272414 A1 WO 2023272414A1
Authority
WO
WIPO (PCT)
Prior art keywords
image
super
resolution
input image
frequency band
Prior art date
Application number
PCT/CN2021/102613
Other languages
English (en)
Chinese (zh)
Inventor
张晓雯
侯耀林
彭剑威
胡伟
周春华
Original Assignee
华为技术有限公司
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 华为技术有限公司 filed Critical 华为技术有限公司
Priority to PCT/CN2021/102613 priority Critical patent/WO2023272414A1/fr
Priority to CN202180041397.7A priority patent/CN115735225A/zh
Publication of WO2023272414A1 publication Critical patent/WO2023272414A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T3/00—Geometric image transformations in the plane of the image
    • G06T3/40—Scaling of whole images or parts thereof, e.g. expanding or contracting

Definitions

  • the present application relates to the field of image technology, and more particularly, to an image processing method and an image processing device.
  • Game super-resolution technology is an important processing technology designed to enhance the resolution and quality of game screens. Different from natural image super-resolution, game image super-resolution is based on rendering model modeling, so the main difference between low-resolution images and high-definition images of natural images lies in details, and the difference between low-resolution images and high-definition images of game screens is not only details Zigzag and outline breakpoints.
  • Embodiments of the present application provide an image processing method and an image processing device, which can improve the processing efficiency of image super-resolution.
  • an image processing method includes: acquiring an input image; dividing the input image to obtain multiple frequency band components of the input image, and the above multiple frequency band components are the same as multiple sub-networks in the super-resolution neural network.
  • One-to-one correspondence determine the number of subnetwork operation layers corresponding to each frequency band component in the multiple frequency band components; use the super-resolution neural network to perform super-resolution processing on the input image, and obtain multiple super-resolution processing results corresponding to the multiple frequency band components;
  • the above multiple super-resolution processing results are synthesized into an output image.
  • the number of operation layers of each sub-network of the super-resolution neural network is mainly determined according to the components of multiple frequency bands of the input image, and then the input image is processed by the super-resolution neural network, thereby ensuring the super-resolution of the image. While improving the accuracy of the results of the resolution processing, it also reduces the occupied computing resources and improves the processing efficiency.
  • the multiple frequency band components include low-frequency components and high-frequency components, and the first loss function of the first sub-network corresponding to the low-frequency component in the multiple sub-networks is different from that of The second loss function of the second sub-network corresponding to the high-frequency component.
  • different loss functions are set for the sub-networks corresponding to different frequency bands, so that the accuracy of the output image is improved.
  • the first loss function is a pixel difference loss function
  • the second loss function is a perceptual difference loss function
  • the image information contained in different frequency band components is different, for example, the above low frequency components include more contour information, and the high frequency components include more detailed information, so in the super-resolution neural network for low frequency components and
  • the computational complexity is different, and the purpose of emphasis is also different.
  • the low-frequency components focus on information such as image jaggies, contour breakpoints, etc., so the pixel difference can be used as the loss function for the low-frequency components, and the high-frequency components focus on Image details and other information, so the perceptual difference can be used as the loss function for the high-frequency component, that is, different reconstruction quality constraints can be set for different frequency bands, so as to achieve better super-resolution processing effects.
  • the input image is any one of a plurality of input blocks obtained by dividing the image to be processed into blocks
  • the above method further includes: combining the plurality of input blocks The corresponding multiple output images are stitched into a target image. And if the input image is the image to be processed, the output image is the target image.
  • block division can reduce the amount of calculation, and on the other hand, it can help to improve the quality of the output image.
  • the image has local correlation, for example, for an image to be processed including sky, river, flowers and plants, if the image to be processed is divided into blocks, the image to be processed can be divided into some images with high internal correlation For example, an input block only includes flowers or only the sky, etc., and makes it more accurate when determining the number of operation layers of the sub-network in the super-resolution neural network.
  • the number of operation layers required by the sub-network in the super-resolution neural network depends more on the most complex content part of the image to be processed, here it is flowers and plants (including more color, border and other details), therefore, more calculation layers are required.
  • the number of computing layers required for each tile is different.
  • the number of computing layers required is more, and for tiles including only the sky, the number of computing layers required The number of layers is less, so the overall accuracy is higher and the amount of calculation is further reduced.
  • the super-resolution neural network may perform super-resolution processing on the feature vector of the input image.
  • wavelet transform when performing frequency division on the input image, wavelet transform may be used to perform frequency division on the input image.
  • the above method further includes: adjusting the size of the target image according to display requirements. That is, processing such as scaling is performed on the target image, and its size is adjusted to meet the requirements of different devices, such as display requirements. Assuming that on a terminal device such as a mobile phone, the size of the target image can be adjusted to adapt to the maximum display resolution supported by the terminal device.
  • a training method for an image processing model comprising: obtaining training data, the training data including low-definition images and labeled high-definition images corresponding to the low-definition images, and updating parameters of the image processing model according to the training data, so that The difference between the labeled high-definition image and the high-definition output image is the smallest.
  • the image processing model includes a depth estimation module and a super-resolution neural network. The multiple frequency band components of the low-resolution image correspond to the sub-networks of the super-resolution neural network.
  • the depth estimation module uses In order to determine the number of sub-network operation layers corresponding to each frequency band component in multiple frequency band components, the super-resolution neural network is used to perform super-resolution processing on low-definition images, and obtain multiple super-resolution processing results corresponding to multiple frequency bands. Image processing The model is also used to synthesize multiple super-resolution processing results into high-definition output images.
  • the image processing model obtained by the training method provided in the second aspect can be used to execute the image processing method in any implementation manner of the first aspect, so as to improve the efficiency and/or image quality of super-resolution processing.
  • the multiple frequency band components include low-frequency components and high-frequency components, and the first loss of the first sub-network corresponding to the low-frequency component among the multiple sub-networks of the super-resolution neural network The function is different from the second loss function of the second sub-network corresponding to the high frequency component.
  • the first loss function is a pixel difference loss function
  • the second loss function is a perceptual difference loss function
  • the super-resolution neural network is specifically used to perform super-resolution processing on the feature vector of the input image according to the number of operation layers, to obtain the above-mentioned super-resolution processing results of the multiple frequency bands.
  • the above-mentioned multiple frequency band components of the low-definition image are obtained by frequency division of the low-definition image by wavelet transform.
  • an image processing device in a third aspect, includes a unit for executing the method in any one implementation manner of the first aspect above.
  • an image processing model training device includes a unit for executing the training method in any one of the above-mentioned implementation manners of the second aspect.
  • an image processing device which includes: a memory for storing a program; a processor for executing the program stored in the memory, and when the program stored in the memory is executed, the processor uses To execute the method in any one of the implementation manners in the first aspect.
  • an image processing model training device includes: a memory for storing a program; a processor for executing the program stored in the memory, and when the program stored in the memory is executed, The processor is configured to execute the training method in any one implementation manner in the second aspect.
  • a computer-readable medium stores program code for execution by a device, where the program code includes a method for executing any one of the implementation manners in the first aspect.
  • a computer-readable medium stores program code for execution by a device, where the program code includes a training method for executing any one of the implementation manners in the second aspect.
  • a computer program product including instructions is provided, and when the computer program product is run on a computer, the computer is made to execute the method in any one of the implementation manners in the first aspect above.
  • a computer program product containing instructions is provided, and when the computer program product is run on a computer, the computer is made to execute the training method in any one of the above-mentioned implementation manners of the second aspect.
  • a chip in an eleventh aspect, includes a processor and a data interface, and the processor reads instructions stored on the memory through the data interface, and executes any one of the implementations in the first aspect above. Methods.
  • the chip may further include a memory, the memory stores instructions, the processor is configured to execute the instructions stored in the memory, and when the instructions are executed, the The processor is configured to execute the method in any one of the implementation manners in the first aspect.
  • a chip in a twelfth aspect, includes a processor and a data interface, the processor reads instructions stored on the memory through the data interface, and executes any one of the implementations in the second aspect above training method.
  • the chip may further include a memory, the memory stores instructions, the processor is configured to execute the instructions stored in the memory, and when the instructions are executed, the The processor is configured to execute the training method in any one of the implementation manners in the second aspect.
  • FIG. 1 is a schematic structural diagram of an image processing device according to an embodiment of the present application.
  • FIG. 2 is a structural diagram of a super-resolution processing unit according to an embodiment of the present application.
  • Fig. 3 is a schematic diagram of the super-resolution processing process of the embodiment of the present application.
  • Fig. 4 is a schematic flowchart of an image processing method according to an embodiment of the present application.
  • Fig. 5 is a schematic flowchart of a method for training an image processing model according to an embodiment of the present application.
  • Fig. 6 is a schematic diagram of an image to be processed and an input block according to an embodiment of the present application.
  • Fig. 7 is a schematic diagram of an output block, a target image and a resized target image according to an embodiment of the present application.
  • FIG. 8 is a schematic diagram of an image processing process in an applicable scenario according to an embodiment of the present application.
  • FIG. 9 is a schematic block diagram of an image processing device according to an embodiment of the present application.
  • FIG. 10 is a schematic diagram of a hardware structure of an image processing device according to an embodiment of the present application.
  • Fig. 11 is a schematic block diagram of an image processing model training device according to an embodiment of the present application.
  • FIG. 12 is a schematic diagram of a hardware structure of an image processing model training device according to an embodiment of the present application.
  • FIG. 1 is a schematic structural diagram of an image processing device according to an embodiment of the present application.
  • the image processing apparatus 1000 may include part or all of the following: a super-resolution processing unit 100 , a blocking unit 200 , a splicing unit 300 and a size adjustment unit 400 . That is, in practice, each unit in FIG. 1 does not necessarily exist.
  • the image processing device 1000 may only include the super-resolution processing unit 100, and for example, the image processing device 1000 may include the super-resolution processing unit 100, the blocking unit 200 and the The stitching unit 300, and for example, the image processing device 1000 may include the super-resolution processing unit 100, the blocking unit 200, the stitching unit 300, and the size adjustment unit 400, and for example, the image processing device 1000 may include the super-resolution processing unit 100 and the size adjustment unit 400 wait.
  • some or all of the aforementioned blocking unit 200 , stitching unit 300 , and resizing unit 400 may also be integrated into the super-resolution processing unit 100 .
  • the image processing apparatus 1000 may also be in product forms such as chips, integrated circuits, and software development kits (software development kit, SDK).
  • the super-resolution processing unit 100 is configured to process a low-definition input image to obtain a high-definition output image, and the input image may be the image to be processed or a sub-block of the image to be processed.
  • the input image is a low-definition (i.e. low-resolution) image
  • the output image is a high-definition (i.e. high-resolution) image obtained after the super-resolution processing unit 100 is used to perform super-resolution processing (hereinafter referred to as super-resolution processing) on the input image. image.
  • the solutions in the embodiments of the present application are especially suitable for artificially synthesized (made) image scenes such as game rendering scenes, such as games, animations, animations, etc., because artificially synthesized images are more prone to images that are too smooth, jagged, Phenomena such as mosaics and breakpoints, but the scheme of the embodiment of the present application can still be applied to natural images (that is, non-artificially synthesized images). Therefore, the image to be processed can be a game rendering image, an animation image frame, an animation image frame, a natural Images, etc., there is no limit.
  • the low-definition input image may be a low-definition game rendering image, a low-definition animation image frame, a low-definition animation image frame, a low-definition natural image, etc., without limitation.
  • the super-resolution processing apparatus 100 is mainly used for performing super-resolution processing on an input image, that is, converting a low-definition image into a high-definition image.
  • converting a low-definition image to a high-definition image the number of image pixels increases, and how to fill these increased pixels so that the conversion of a low-definition image into a high-definition image can be understood as the so-called super-resolution processing (super-resolution processing) .
  • super-resolution processing may include traditional super-resolution processing methods and neural network-based super-resolution processing methods.
  • the traditional super-resolution processing method is based on pixel interpolation to improve the resolution of the image, that is, on the basis of the original image pixels, a new element is inserted between the pixel point values using an appropriate interpolation algorithm.
  • Common interpolation schemes include linear interpolation, bilinear interpolation, nearest neighbor interpolation, bicubic interpolation, etc.
  • the calculation of the surrounding pixels supplements the missing pixel values, which cannot solve the problems of the connection of contour breakpoints in the image and the elimination of jaggies.
  • the reason for the above phenomenon in the existing neural network-based super-resolution processing method is that the image processing is "one size fits all" type, that is, all images are exactly the same processing process and processing parameters, while In fact, even if the size is the same between images, the richness of the contained content is different, which can also be understood as the details included in the images are different. For images with rich contours, more attention should be paid to the extraction of contour information, and for images with rich details, more attention should be paid to the extraction of detail information.
  • the existing technology does not consider the impact of the difference in content richness, so there is no way to simultaneously Ensuring that both types of information can be effectively extracted will lead to poor ability to extract outline information, or poor ability to extract detail information, or both.
  • the number of operation layers of each sub-network of the super-resolution neural network is mainly determined according to the components of multiple frequency bands of the input image, and then the feature vector of the input image of the super-resolution neural network is used for processing, thereby While ensuring the accuracy of image super-resolution processing results, it also reduces the occupied computing resources and improves processing efficiency.
  • the specific content will be introduced below and will not be repeated here.
  • the blocking unit 200 is configured to block the image to be processed (low-definition) to obtain multiple input image blocks (low-definition).
  • the input tiles can be input into the super-resolution processing unit 100 to obtain multiple output tiles (high-definition).
  • the splicing unit 300 is used to splice multiple input image blocks to obtain a target image (high-definition).
  • the size adjustment unit 400 is used to adjust the size of the target image to meet the display requirements, that is, to obtain a display image that meets the display requirements.
  • the display image may be, for example, a presentation image corresponding to the aforementioned game rendering image on the terminal device.
  • FIG. 2 is a structural diagram of a super-resolution processing unit according to an embodiment of the present application.
  • a high-definition output image can be obtained by inputting a low-definition input image into the super-resolution processing unit 100 .
  • the input image is a low-definition (i.e. low-resolution) image
  • the output image is a high-definition (i.e. high-resolution) image obtained after the super-resolution processing unit 100 is used to perform super-resolution processing (hereinafter referred to as super-resolution processing) on the input image. image.
  • super-resolution processing super-resolution processing
  • FIG. 2 can be regarded as an example of the case where the image processing apparatus 1000 only includes the super-resolution processing unit 100 , and can also be regarded as an example of the super-resolution processing unit 100 in the image processing apparatus 1000 of other structural forms.
  • the super-resolution processing unit 100 includes a feature extraction module 110 , a frequency division module 120 , a depth estimation module 130 , a super-resolution neural network 140 and an image reconstruction module 150 .
  • the feature extraction module 110 is used to perform feature extraction on the input image to obtain a feature vector of the input image.
  • the frequency division module 120 is used to divide the frequency of the input image to obtain multiple frequency components of the input image.
  • the low frequency component and high frequency component of the input image can be obtained.
  • the frequency division method can adopt wavelet transform, then the input image can be processed by wavelet transform, and the component maps of four input images of LL, LH, HL, and HH can be obtained, wherein L represents low (low), and H represents high (high) , that is, the components of the four frequency bands of the input image are obtained after wavelet transformation.
  • the depth estimation module 130 is used to determine the number of computing layers required for each frequency band component in the multiple frequency band components according to the computing requirements of each frequency band component in the multiple frequency band components.
  • the depth estimation module 130 can use machine learning or neural network methods to obtain a model of the corresponding relationship between frequency band components and the number of calculation layers, and then use this model to obtain the number of calculation layers required by the multiple frequency band components of the input image.
  • it can be realized by using a convolutional neural network (CNN) model, which is equivalent to that the input data of the CNN model is multiple frequency band components, and the output data is the number of computing layers corresponding to the multiple frequency band components.
  • CNN convolutional neural network
  • the super-resolution neural network 140 includes a plurality of sub-networks, for example, a sub-network 141 to a sub-network 144 are shown in FIG. 1 , and the sub-networks correspond one-to-one to the above-mentioned multiple frequency band components.
  • the super-resolution neural network 140 is used to perform super-resolution processing on the feature vectors obtained by the feature extraction module 110 to obtain super-resolution processing results of multiple frequency bands.
  • the output of the sub-band module 120 is the component maps of the four input images of LL, LH, HL, and HH
  • the sub-network 141 to sub-network 144 respectively processes the feature vectors input by the feature extraction module 110 with the number of operation layers to obtain 4 super-resolution processing results (for example, 4 super-resolution processing result maps).
  • the image information contained in different frequency band components is different, for example, the low-frequency components mentioned above include more contour information, and the high-frequency components include more detailed information.
  • the computational complexity is different, and the purpose of emphasis is also different.
  • the low-frequency components focus on information such as image jaggies, contour breakpoints, etc., so the pixel difference can be used as the loss function for the low-frequency components
  • the high-frequency components Focus on the details of the image and other information, so the perceptual difference can be used as the loss function for the high-frequency component, that is, different reconstruction quality constraints can be set for different frequency bands, so as to achieve better super-resolution processing effects.
  • the component maps of the four input images of LL, LH, HL, and HH can correspond to the pixel difference loss function for the LL component, and correspond to the perceptual difference loss function for the LH, HL, and HH components.
  • the image reconstruction module 150 is used to synthesize the super-resolution processing results of multiple frequency bands obtained by the above-mentioned super-resolution neural network 140 into a high-definition output image. That is, the inverse transform of frequency division is used. If the frequency division adopts wavelet transform, then here is the method of using inverse wavelet transform (or called wavelet upsampling convolution).
  • the wavelet upsampling convolution operator is different from the traditional upsampling convolution operator in structure, and it is an inverse operator specially for wavelet transform.
  • the super-resolution processing unit 100 shown in FIG. 2 mainly divides the frequency of the input image, and then, when processing different frequency bands, different sub-network calculation layers can be set for different frequency bands, so that the waste of computing resources can be reduced. Thus, the computational efficiency is improved. In addition, better image super-resolution processing effects can be achieved.
  • the low-frequency component uses the pixel difference as the loss function
  • the high-frequency part uses the perceptual difference as the loss function, so it can effectively solve the problem that the output image in the prior art is too smooth, jagged, and mosaic. , breakpoints and other issues.
  • the image to be processed can also be divided into blocks to obtain multiple input tiles of the image to be processed, that is, the above input image is Process any one of the multiple input blocks of the image, and then input these input blocks to the super-resolution processing unit 100 respectively to obtain a high-definition output image (also referred to as an output image) corresponding to the multiple input blocks. blocks), and then these high-definition output tiles can be stitched into the target image. That is, multiple input tiles of the image to be processed are sequentially input to the super-resolution processing unit 100 to obtain multiple high-definition output tiles, and then the multiple high-definition output tiles are spliced into a target image.
  • the above input image is Process any one of the multiple input blocks of the image, and then input these input blocks to the super-resolution processing unit 100 respectively to obtain a high-definition output image (also referred to as an output image) corresponding to the multiple input blocks. blocks), and then these high-definition output tiles can be stitched into the target image
  • block division can reduce the amount of calculation, and on the other hand, it can help to improve the quality of the output image.
  • the image has local correlation, for example, for an image to be processed including sky, river, flowers and plants, if the image to be processed is divided into blocks, the image to be processed can be divided into some images with high internal correlation For example, an input block only includes flowers or only the sky, etc., and makes the subsequent determination of the number of operation layers of the sub-network in the super-resolution neural network 140 more accurate.
  • the number of computing layers required by the sub-network in the super-resolution neural network 140 depends more on the most complex content part in the image to be processed, here it is flowers and plants (including more Many colors, borders and other details), so more calculation layers are required.
  • the number of computing layers required for each tile is different.
  • the number of computing layers required is more, and for tiles including only the sky, the number of computing layers required The number of layers is less, so the overall accuracy is higher and the amount of calculation is further reduced.
  • the input image is the image to be processed itself, and the output image is the target image, that is, the input image is an image to be processed that has not been segmented, and the output image does not need to be spliced. That is to say, the input image may be a complete image to be processed, or an input block after the image to be processed is divided into blocks.
  • the target image is its corresponding high-definition output image; when the input image is a block of the image to be processed, the target image is spliced from the output images.
  • Fig. 3 is a schematic diagram of the super-resolution processing process of the embodiment of the present application.
  • the input image in FIG. 3 may be a complete image to be processed, or an input block after the image to be processed is divided into blocks.
  • the output image is the target image; when the input image is a block of the image to be processed, the output image is a high-definition output block, and splicing is required to obtain the target image.
  • the input image is a low-resolution image, and the output image is a high-definition image.
  • the input image is carried out to wavelet transform (indicated by DWT in the figure), to obtain four frequency band components of the input image, this part can be regarded as the frequency division module 120 divides the input image to obtain the frequency of the input image
  • wavelet transform indicated by DWT in the figure
  • Convolute the four frequency band components (the convolutional neural network is represented by CNN in the figure) to obtain the estimated value of the number of operation layers (also called the estimated depth value, which can be represented by a depth map).
  • This part can be regarded as a depth estimate Module 130 determines a specific example of the number of operation layers of each sub-network of the super-resolution neural network according to multiple frequency band components.
  • this part can be regarded as a specific example of feature extraction module 110 performing feature extraction on the input image to obtain the feature vector of the input image .
  • both the feature vector and the number of operation layers are input to multiple sub-networks (that is, sub-network #1 to sub-network #4), and the high-definition super-resolution processing results corresponding to multiple frequency band components are obtained (LL', LH', HL ', HH'), this part can be regarded as a specific example in which the super-resolution neural network 140 processes the feature vectors obtained by the above-mentioned feature extraction module 110 to obtain the super-resolution processing results of multiple frequency bands.
  • Inverse wavelet transform (indicated by IDWT in the figure) is performed on multiple high-definition super-resolution processing results to obtain high-definition output images.
  • further processing such as scaling may be performed on the target image, and its size may be adjusted to meet requirements of different devices, such as display requirements. Assuming that on a terminal device such as a mobile phone, the size of the target image can be adjusted to adapt to the maximum display resolution supported by the terminal device.
  • the training data can be the low-definition image and the labeled high-definition image corresponding to the low-definition image.
  • Labeled high-definition images can be understood as real high-definition images.
  • a real high-definition image can be compressed into a low-resolution image, so that the low-resolution image can be used as training data, and the labeled high-definition image can be used as the label data of the training data to update the depth Estimation module, parameters of super-resolution neural network.
  • the loss function of subnetwork #1 can be the pixel difference loss function (or It can be called the objective quality (object quality))
  • the loss function of subnetwork #2 to subnetwork #4 can be the perceptual difference loss function (or called the perceptual quality (pereceptual quality) function.
  • Figure 3 shows the label HD image, and the wavelet transform of the label HD image to obtain four frequency band components: LL, LH, HL, HH.
  • the loss function of subnetwork #1 is to minimize the pixel difference between LL and LL', which can be understood as improving The ability of the sub-band neural network to extract contour information; the loss function of sub-network #2 to sub-network #4 is to minimize the difference in perceptual quality between LL and LL', which can be understood as improving the ability of the sub-band neural network to extract detailed information.
  • FIG. 3 also shows waveform diagrams corresponding to LL, LH, HL, and HH respectively, so as to facilitate understanding of frequency band components.
  • Fig. 4 is a schematic flowchart of an image processing method according to an embodiment of the present application. Each step in FIG. 4 will be introduced below.
  • 401. Acquire an input image.
  • the input image may be a low-definition game rendering image, a low-definition animation image frame, a low-definition animation image frame, a low-definition natural image, etc., without limitation.
  • a low-definition input image can be obtained by using a camera, a camera, etc., and the input image can also be further processed by an image signal processor (ISP), or the input image can be read from a storage device, or can be read through a communication interface. etc. Get the input image from the network.
  • the input image may be an image to be processed, or any block among multiple subdivided blocks of the image to be processed.
  • Step 402 Perform frequency division on the input image to obtain multiple frequency band components of the input image, where the multiple frequency band components correspond to multiple sub-networks in the super-resolution neural network one by one.
  • Step 402 can be performed by using the frequency division module described above, and the frequency division method using wavelet transform in FIG. 3 can also be referred to, which will not be repeated here for brevity.
  • Multiple frequency band components may include low-frequency components and high-frequency components, and related introductions may also refer to relevant content above.
  • the number of subnetwork operation layers corresponding to each frequency band component in the multiple frequency band components can be determined according to the computing requirements of each frequency band component in the multiple frequency band components.
  • Computational requirements can be understood as, for different frequency band components, the computational complexity is different, so the required network depth is also different.
  • the super-resolution neural network may include multiple sub-networks, such as a first sub-network and a second sub-network, wherein the first sub-network is used to perform super-resolution processing on low-frequency components in multiple frequency band components, and the second sub-network
  • the sub-network is used to perform super-resolution processing on high-frequency components in multiple frequency band components. That is, the first subnetwork is a subnetwork corresponding to low frequency components, and the second subnetwork is a subnetwork corresponding to high frequency components.
  • the first subnetwork and the second subnetwork are different networks among the plurality of subnetworks, and the first loss function of the first subnetwork is different from the second loss function of the second subnetwork.
  • the first sub-network focuses on objective quality (pixel difference), that is, the first loss function is a pixel difference loss function; the second sub-network focuses on perceptual quality (perceptual difference), and the second loss function is a perceptual difference loss function.
  • the super-resolution neural network can refer to the above introduction, so I won’t go into details.
  • the super-resolution neural network can perform super-resolution processing on the input image, or it can perform super-resolution processing on the feature vector of the input image.
  • the feature vector of the input image may be obtained by feature extraction of the input image, for example, the feature vector may be extracted by using a feature extraction network, which may be realized by using the feature extraction module described above.
  • Step 405 Synthesize the above multiple super-resolution processing results into an output image.
  • Step 405 can be performed by using the above-mentioned image reconstruction module.
  • the method shown in Figure 4 mainly determines the number of operation layers of each sub-network of the super-resolution neural network according to the components of multiple frequency bands of the input image, and then uses the super-resolution neural network to input the image for processing, thus ensuring image super-resolution While the processing results are accurate, the occupied computing resources are also reduced, and the processing efficiency is improved. Then, by taking into account the different characteristics of the image in different frequency bands, different loss functions are set for the sub-networks corresponding to different frequency bands, so that the accuracy of the output image is improved.
  • the above-mentioned input image may also be any one of multiple input blocks obtained by dividing the image to be processed into blocks, and multiple output images corresponding to the multiple input blocks need to be spliced into the target image. If the input image is the image to be processed, the output image is the target image. Blocking can greatly reduce the amount of computation and further improve computing efficiency. After block, based on the local correlation of the image, the computing requirements are more accurate, and the determination of the number of computing layers is also more accurate, thereby improving the quality of the output image.
  • the size of the target image can also be adjusted according to display requirements.
  • the target image can be further processed by scaling and other processing, and its size can be adjusted to meet the requirements of different devices, such as display requirements. Assuming that on terminal devices such as mobile phones, computers, and smart screens, the size of the target image can be adjusted to adapt to the maximum display resolution supported by the terminal device.
  • Fig. 5 is a schematic flowchart of a method for training an image processing model according to an embodiment of the present application. Each step in FIG. 5 will be introduced below.
  • 501. Obtain training data.
  • the training data consists of low-resolution images and labeled high-resolution images corresponding to the low-resolution images.
  • the tagged high-definition image is the real high-definition image of the low-definition image, which plays the role of a label. It is equivalent to the goal of the training phase to make the processing result of the low-definition image closer to the tagged high-definition image, the better, or understood as the difference between the tagged high-definition image and the low-definition image. The smaller the difference in processing results for clear images, the better.
  • the low-definition image can be used as the above-mentioned low-definition image
  • the high-definition image can be used as the above-mentioned tagged high-definition image.
  • the above images can also be tiles, that is, small tiles cut out from a large image, then the low-definition image is a low-definition tile, and the tagged high-definition image is the tagged high-definition corresponding to the low-definition tile. tiles. For a block, it can also be regarded as an image with a relatively small size, so the introduction will not be repeated.
  • the image processing model may include a depth estimation module and a super-resolution neural network.
  • the depth estimation module is used to determine the number of sub-network operation layers corresponding to each frequency band component in a plurality of frequency band components, and the super-resolution neural network is used to perform super-resolution on low-resolution images.
  • the image processing model is also used to synthesize the super-resolution processing results of multiple frequency bands into a high-definition output image.
  • the depth estimation module may be a convolutional neural network.
  • the depth estimation module and the super-resolution neural network can be the depth estimation module and the super-resolution neural network as shown in FIG. 2 , and reference can be made to the related introductions in FIG. 2 and FIG. 3 .
  • the image processing model can also include at least one of the following modules: a frequency division module, an image reconstruction module, a feature extraction module, etc., and the above-mentioned modules can also be corresponding modules as shown in Figure 2, which can be related to Figure 2 and Figure 3 introduce.
  • the above-mentioned multiple frequency band components may include low frequency components and high frequency components
  • the first loss function of the first subnetwork corresponding to the low frequency component in the super-resolution neural network may be different from the second subnetwork corresponding to the high frequency component
  • the loss function of the sub-network corresponding to the low-frequency component ie, the first loss function
  • the loss function of the sub-network corresponding to the high-frequency component ie, the second loss function
  • the perceptual For the difference loss function refer to the relevant introductions in step 402 and step 403, and details will not be repeated here.
  • the super-resolution neural network may be specifically used to perform super-resolution processing on feature vectors according to the number of operation layers, to obtain super-resolution processing results of multiple frequency bands.
  • the feature vector can be obtained by using the feature extraction module described above, and will not be described again.
  • the image to be processed is a low-definition large-size image
  • the input block is a low-definition small-size image cut out from the image to be processed, as shown in Figure 6, and (a) in Figure 6 is The image to be processed, (b) in Figure 6 is the input block.
  • the output tile is a small-sized high-definition tile corresponding to the input tile
  • the target image is a large-sized high-definition image spliced by the output tiles, that is, the high-definition large-size image corresponding to the image to be processed.
  • the image output by the size adjustment unit 400 is the image after the target image is resized, as shown in FIG. 7 , (a) in FIG. 7 is a high-definition output block, and (b) in FIG. 7 is the target image. (c) in Figure 7 is the resized target image.
  • FIG. 8 is a schematic diagram of an image processing process in an applicable scenario according to an embodiment of the present application.
  • the terminal device may be various devices capable of displaying images, such as a mobile phone, a computer, a smart screen, and a game device.
  • the super-resolution processing system can be any image processing device 1000 shown in FIG. 1, but it should be understood that FIG. 8 is only a specific example. In the actual process, the specific interaction process can also be other situations, as long as the embodiment of the present application can be applied image processing methods. However, it should be understood that the super-resolution processing system can be integrated in the terminal device, or it can be a cloud device, server, host, etc. When the super-resolution processing system is integrated in the terminal device, steps 801-803 and step 808 are performed by the terminal device The internal circuit is realized and does not need to communicate with the outside through the communication interface.
  • the super-resolution processing system can also be deployed on terminal devices and data processing devices (such as the above-mentioned cloud devices, servers, hosts, etc.), for example, the part deployed on the terminal device side is used to perform steps 804, 806 and 807, the part deployed on the data processing device side is used to perform step 805; for another example, the part deployed on the terminal device side is used to perform steps 804 and 807, and the part deployed on the data processing device side is used to perform steps 805 and 806; For another example, the part deployed on the terminal device side is used to execute step 804, and the part deployed on the data processing device side is used to execute steps 805-807.
  • This situation requires data interaction between the terminal device and the data processing device.
  • the terminal device needs to send the input image (the image to be processed or the block of the image to be processed) to the data processing device, and the data processing device needs to send the output The image or target image or display image is sent to the terminal device.
  • Step 801 is mainly used to confirm that the super-resolution processing system can support the super-resolution processing of the terminal device, for example, by judging whether the super-resolution processing system supports the model of the terminal device, that is, whether the terminal device is included in the list of supported models models.
  • the super-resolution processing system returns support status to the terminal device. That is to say, the super-resolution processing system informs the terminal device whether it supports it. If it is supported, the following steps 803-808 can be performed; if not, the terminal device can use a traditional method to perform subsequent processing. No matter whether it is supported or not, the terminal device can be notified, and if not supported, the subsequent steps will not be performed.
  • the terminal device sends the image to be processed to the super-resolution processing system.
  • the image to be processed can be a synthetic image or a natural image.
  • the super-resolution processing system divides the image to be processed into blocks to obtain an input image.
  • the input image is any one of the block blocks of the image to be processed, you can refer to the introduction above. However, it should be understood that subsequent processing may also be directly performed on the entire image to be processed, and step 804 is not required at this time.
  • the super-resolution processing system processes the input image to obtain an output image.
  • the output image and the above processing can refer to the introduction above. If the entire image to be processed is processed directly, the input image is the image to be processed, and the output image is the target image. 806.
  • the super-resolution processing system stitches the output images into a target image.
  • Step 806 may or may not be executed.
  • step 806 needs to be executed, but when step 804 is not executed, step 806 does not need to be executed.
  • the super-resolution processing system adjusts the target image size.
  • Step 807 is mainly to make the image more suitable for display requirements of the terminal device. Therefore, if there is no such requirement, step 807 may not be executed.
  • the super-resolution processing system returns the display image to the terminal device.
  • the display image may be the target image, or the target image after resizing, that is to say, when step 807 is executed, the resized target image is the display image, and when step 807 is not executed, the target image is the display image .
  • Fig. 8 is only an example of an applicable scenario of the solution of the embodiment of the present application, and the steps included in Fig. 8 may also exist in various situations, as long as the solution of the embodiment of the present application can be realized Any combination and order is fine.
  • the image processing apparatus will be introduced below with reference to FIG. 9 .
  • the image processing device shown in FIG. 9 can be used to execute various steps of the image processing method of the embodiment of the present application, and the image processing device can be a computer, server, or other device with sufficient computing power to perform image super-resolution processing.
  • FIG. 9 is a schematic block diagram of an image processing device according to an embodiment of the present application.
  • the apparatus 2000 shown in FIG. 9 includes an acquisition unit 2001 and a processing unit 2002 .
  • the apparatus 2000 may be used to execute the steps of the image processing method of the embodiment of the present application.
  • the acquiring unit 2001 may be used to execute step 401 in the method shown in FIG. 4
  • the processing unit 2002 may be used to execute step 402 to step 405 in the method shown in FIG. 4 .
  • the device 2000 may also be the super-resolution processing system shown in FIG. 8 , or have the functions of the super-resolution processing system shown in FIG. 8 .
  • the acquiring unit 2001 may be equivalent to the communication interface 3003 in the device 3000 shown in FIG.
  • the processor 3002 in the shown apparatus 3000 can obtain the above-mentioned input image from the memory 3001 through the processor 3002 at this time.
  • the processing unit 2002 in the apparatus 2000 shown in FIG. 9 may be equivalent to the processor 3002 in the apparatus 3000 shown in FIG. 10 .
  • FIG. 10 is a schematic diagram of a hardware structure of an image processing device according to an embodiment of the present application.
  • the device 3000 shown in FIG. 10 includes a memory 3001 , a processor 3002 , a communication interface 3003 and a bus 3004 .
  • the memory 3001 , the processor 3002 , and the communication interface 3003 are connected to each other through a bus 3004 .
  • the memory 3001 may include a read only memory (read only memory, ROM), a static storage device, a dynamic storage device or a random access memory (random access memory, RAM).
  • the memory 3001 may store programs, and when the programs stored in the memory 3001 are executed by the processor 3002, the processor 3002 and the communication interface 3003 are used to execute various steps of the image processing method of the embodiment of the present application.
  • Processor 3002 may include a general-purpose central processing unit (central processing unit, CPU), a microprocessor, an application-specific integrated circuit (application specific integrated circuit, ASIC), a graphics processing unit (graphics processing unit, GPU) or one or more An integrated circuit is used to execute related programs to realize the functions required by the units in the image processing device of the embodiment of the present application, or to execute various steps of the image processing method of the embodiment of the present application.
  • CPU central processing unit
  • ASIC application specific integrated circuit
  • GPU graphics processing unit
  • An integrated circuit is used to execute related programs to realize the functions required by the units in the image processing device of the embodiment of the present application, or to execute various steps of the image processing method of the embodiment of the present application.
  • the processor 3002 may also be an integrated circuit chip with signal processing capabilities. During implementation, each step of the image processing method in the embodiment of the present application may be completed by an integrated logic circuit of hardware in the processor 3002 or instructions in the form of software.
  • the above-mentioned processor 3002 may also include a general-purpose processor, a digital signal processor (digital signal processing, DSP), an ASIC, a field programmable gate array (field programmable gate array, FPGA) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.
  • DSP digital signal processing
  • ASIC application specific integrated circuit
  • FPGA field programmable gate array
  • Various methods, steps, and logic block diagrams disclosed in the embodiments of the present application may be implemented or executed.
  • a general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.
  • the steps of the image processing method disclosed in the embodiments of the present application may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor.
  • the software module can be located in a mature storage medium in the field such as random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, register.
  • the storage medium is located in the memory 3001, and the processor 3002 reads the information in the memory 3001, and combines its hardware to complete the functions required by the units included in the image processing device of the embodiment of the present application, or execute the image processing method of the embodiment of the present application each step.
  • the communication interface 3003 implements communication between the apparatus 3000 and other devices or communication networks by using a transceiver device such as but not limited to a transceiver.
  • a transceiver device such as but not limited to a transceiver.
  • the target image can be transmitted through the communication interface 3003 .
  • the bus 3004 may include a pathway for transferring information between various components of the device 3000 (eg, memory 3001 , processor 3002 , communication interface 3003 ).
  • Fig. 11 is a schematic block diagram of an image processing model training device according to an embodiment of the present application.
  • the image processing model training apparatus 4000 shown in FIG. 11 includes an acquisition unit 4001 and a training unit 4002 .
  • the acquisition unit 4001 and the training unit 4002 can be used to execute the training method of the living body detection model in the embodiment of the present application.
  • the acquisition unit 4001 can perform the above step 401
  • the training unit 4002 can perform the above step 402.
  • the training unit 4002 in the above device 4000 may be equivalent to the processor 5002 in the device 5000 hereinafter.
  • FIG. 12 is a schematic diagram of a hardware structure of an image processing model training device according to an embodiment of the present application.
  • the training device 5000 shown in FIG. 12 includes a memory 5001 , a processor 5002 , a communication interface 5003 and a bus 5004 .
  • the memory 5001 , the processor 5002 , and the communication interface 5003 are connected to each other through a bus 5004 .
  • the memory 5001 may include ROM, static storage, dynamic storage or RAM.
  • the memory 5001 can store programs, and when the programs stored in the memory 5001 are executed by the processor 5002, the processor 5002 and the communication interface 5003 are used to execute each step of the image processing model training method of the embodiment of the present application.
  • the processor 5002 may include a CPU, a microprocessor, an ASIC, a GPU, or one or more integrated circuits, and is used to execute related programs, so as to realize the functions required by the units in the image processing model training device of the embodiment of the present application, Or execute the image processing model training method of the method embodiment of the present application.
  • the processor 5002 may also be an integrated circuit chip with signal processing capabilities. During implementation, each step of the image processing model training method of the present application may be completed by an integrated logic circuit of hardware in the processor 5002 or instructions in the form of software.
  • the aforementioned processor 5002 may also include a general-purpose processor, DSP, ASIC, FPGA or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components.
  • a general-purpose processor may be a microprocessor, or the processor may be any conventional processor, or the like.
  • the steps of the methods disclosed in the embodiments of the present application may be directly implemented by a hardware decoding processor, or implemented by a combination of hardware and software modules in the decoding processor.
  • the software module can be located in a mature storage medium in the field such as random access memory, flash memory, read-only memory, programmable read-only memory or electrically erasable programmable memory, register.
  • the storage medium is located in the memory 5001, and the processor 5002 reads the information in the memory 5001, and combines its hardware to complete the functions required by the units included in the image processing model training device of the embodiment of the application, or execute the method embodiment of the application A training method for image processing models.
  • the communication interface 5003 implements communication between the apparatus 5000 and other devices or communication networks by using a transceiver device such as but not limited to a transceiver.
  • a transceiver device such as but not limited to a transceiver.
  • the above-mentioned first training data may be obtained through the communication interface 5003 .
  • the bus 5004 may include a pathway for transferring information between various components of the device 5000 (eg, memory 5001, processor 5002, communication interface 5003).
  • the above-mentioned devices 3000 and 5000 only show memory, processors, and communication interfaces, those skilled in the art should understand that the devices 3000 and 5000 may also include other devices required. Meanwhile, according to specific needs, those skilled in the art should understand that the apparatus 3000 may also include hardware devices for implementing other additional functions. In addition, those skilled in the art should understand that the apparatus 3000 may only include the components required to implement the embodiments of the present application, and does not necessarily include all the components shown in FIG. 10 and FIG. 12 .
  • the embodiment of the present application does not specifically limit the specific structure of the execution subject of the method provided in the embodiment of the present application, as long as the program that records the code of the method provided in the embodiment of the present application can be executed according to the method provided in the embodiment of the present application Just communicate.
  • the subject of execution of the method provided by the embodiment of the present application may be a terminal device or a network device, or a functional module in the terminal device or network device that can call a program and execute the program.
  • Computer-readable media may include, but are not limited to, magnetic storage devices (such as hard disks, floppy disks, or tapes, etc.), optical disks (such as compact discs (compact disc, CD), digital versatile discs (digital versatile disc, DVD), etc. ), smart cards and flash memory devices (for example, erasable programmable read-only memory (EPROM), card, stick or key drive, etc.).
  • magnetic storage devices such as hard disks, floppy disks, or tapes, etc.
  • optical disks such as compact discs (compact disc, CD), digital versatile discs (digital versatile disc, DVD), etc.
  • smart cards and flash memory devices for example, erasable programmable read-only memory (EPROM), card, stick or key drive, etc.
  • the processor includes a general-purpose processor, DSP, ASIC, FPGA or other programmable logic devices, discrete gate or transistor logic devices, and discrete hardware components
  • the memory storage module
  • the memories described herein are intended to include, but are not limited to, these and any other suitable types of memories.
  • the disclosed devices and methods may be implemented in other ways.
  • the device embodiments described above are only illustrative.
  • the division of the units is only a logical function division. In actual implementation, there may be other division methods.
  • multiple units or components can be combined or May be integrated into another system, or some features may be ignored, or not implemented.
  • the mutual coupling or direct coupling or communication connection shown or discussed may be through some interfaces, and the indirect coupling or communication connection of devices or units may be in electrical, mechanical or other forms.
  • the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or may be distributed to multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.
  • each functional unit in each embodiment of the present application may be integrated into one unit, each unit may exist separately physically, or two or more units may be integrated into one unit.
  • the functions described above are realized in the form of software function units and sold or used as independent products, they can be stored in a computer-readable storage medium.
  • the essence of the technical solution of this application, or the part that contributes to the prior art, or the part of the technical solution can be embodied in the form of computer software products, which are stored in a storage
  • the computer software product includes several instructions, which are used to make a computer device (which may be a personal computer, server, or network device, etc.) execute all or part of the steps of the methods described in the various embodiments of the present application.
  • the aforementioned storage medium may include, but is not limited to: various media capable of storing program codes such as U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk.

Landscapes

  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

La présente demande, qui relève du domaine de l'intelligence artificielle, concerne un procédé de traitement d'image et un appareil de traitement d'image. Le procédé comprend : l'acquisition d'une image d'entrée ; la réalisation d'une division de fréquence sur l'image d'entrée pour obtenir une pluralité de composantes de bande de fréquences de l'image d'entrée, la pluralité de composantes de bande de fréquences étant dans une correspondance biunivoque avec une pluralité de sous-réseaux dans un réseau neuronal à super-résolution ; la détermination du nombre de couches opérationnelles des sous-réseaux correspondant à chaque composante de la pluralité de composantes de bande de fréquences ; la réalisation d'un traitement à super-résolution sur l'image d'entrée à l'aide du réseau neuronal à super-résolution pour obtenir une pluralité de résultats de traitement à super-résolution correspondant à la pluralité de composantes de bande de fréquences ; et la synthèse de la pluralité de résultats de traitement à super-résolution en une image de sortie. Selon la solution, le nombre de couches opérationnelles de chaque sous-réseau du réseau neuronal à super-résolution est déterminé principalement selon la pluralité de composantes de bande de fréquences de l'image d'entrée, et ensuite l'image d'entrée est traitée à l'aide du réseau neuronal à super-résolution, ce qui réduit les ressources opérationnelles occupées tout en garantissant la précision des résultats de traitement à super-résolution de l'image, et améliore l'efficacité du traitement.
PCT/CN2021/102613 2021-06-28 2021-06-28 Procédé de traitement d'image et appareil de traitement d'image WO2023272414A1 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
PCT/CN2021/102613 WO2023272414A1 (fr) 2021-06-28 2021-06-28 Procédé de traitement d'image et appareil de traitement d'image
CN202180041397.7A CN115735225A (zh) 2021-06-28 2021-06-28 图像处理方法和图像处理装置

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2021/102613 WO2023272414A1 (fr) 2021-06-28 2021-06-28 Procédé de traitement d'image et appareil de traitement d'image

Publications (1)

Publication Number Publication Date
WO2023272414A1 true WO2023272414A1 (fr) 2023-01-05

Family

ID=84690897

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2021/102613 WO2023272414A1 (fr) 2021-06-28 2021-06-28 Procédé de traitement d'image et appareil de traitement d'image

Country Status (2)

Country Link
CN (1) CN115735225A (fr)
WO (1) WO2023272414A1 (fr)

Families Citing this family (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN117726543B (zh) * 2023-07-14 2024-08-27 荣耀终端有限公司 图像处理方法和装置
CN118570609A (zh) * 2024-07-11 2024-08-30 重庆酷贝科技发展有限公司 一种图像识别方法、装置、设备以及存储介质

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108805814A (zh) * 2018-06-07 2018-11-13 西安电子科技大学 基于多频段深度卷积神经网络的图像超分辨重建方法
CN110599401A (zh) * 2019-08-19 2019-12-20 中国科学院电子学研究所 遥感图像超分辨率重建方法、处理装置及可读存储介质
CN111328448A (zh) * 2017-12-01 2020-06-23 华为技术有限公司 图像处理的方法和设备
CN112446436A (zh) * 2020-12-11 2021-03-05 浙江大学 基于生成对抗网络的抗模糊无人驾驶车辆多目标跟踪方法

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN111328448A (zh) * 2017-12-01 2020-06-23 华为技术有限公司 图像处理的方法和设备
CN108805814A (zh) * 2018-06-07 2018-11-13 西安电子科技大学 基于多频段深度卷积神经网络的图像超分辨重建方法
CN110599401A (zh) * 2019-08-19 2019-12-20 中国科学院电子学研究所 遥感图像超分辨率重建方法、处理装置及可读存储介质
CN112446436A (zh) * 2020-12-11 2021-03-05 浙江大学 基于生成对抗网络的抗模糊无人驾驶车辆多目标跟踪方法

Also Published As

Publication number Publication date
CN115735225A (zh) 2023-03-03

Similar Documents

Publication Publication Date Title
Karras et al. Alias-free generative adversarial networks
CN113034358B (zh) 一种超分辨率图像处理方法以及相关装置
EP4030379A1 (fr) Procédé de traitement d'image, dispositif intelligent, et support de stockage lisible par ordinateur
JP6961139B2 (ja) 知覚的な縮小方法を用いて画像を縮小するための画像処理システム
JP7482253B2 (ja) 画像処理方法、装置、コンピュータ機器及び記憶媒体
Chang et al. Vornet: Spatio-temporally consistent video inpainting for object removal
WO2023272414A1 (fr) Procédé de traitement d'image et appareil de traitement d'image
US7499108B2 (en) Image synthesis apparatus, electrical apparatus, image synthesis method, control program and computer-readable recording medium
US11645737B2 (en) Skin map-aided skin smoothing of images using a bilateral filter
CN109996023A (zh) 图像处理方法和装置
JP2023538020A (ja) オブジェクトカテゴリー分類に基づく画像処理
CN103440674A (zh) 一种数字图像蜡笔特效的快速生成方法
CN115330986B (zh) 一种分块渲染模式图形处理方法及系统
CN113724151B (zh) 一种图像增强方法、电子设备和计算机可读存储介质
CN115797176A (zh) 一种图像超分辨率重建方法
CN102611842B (zh) 图像处理设备和图像处理方法
CN114782460A (zh) 图像分割模型的生成方法及图像的分割方法、计算机设备
WO2024032331A9 (fr) Procédé et appareil de traitement d'image, dispositif électronique et support de stockage
Hu et al. Hybrid shift map for video retargeting
CN112200719A (zh) 图像处理方法、电子设备及可读存储介质
Song et al. Multi-curve translator for high-resolution photorealistic image translation
US12094079B2 (en) Reference-based super-resolution for image and video enhancement
Gong et al. Attention-guided network with hierarchical global priors for low-light image enhancement
Xu et al. Joint learning of super-resolution and perceptual image enhancement for single image
Gao et al. A Low-Complexity End-to-End Stereo Matching Pipeline From Raw Bayer Pattern Images to Disparity Maps

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: 21947380

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 21947380

Country of ref document: EP

Kind code of ref document: A1